The biggest news

• Star Citizen telemetry stays on unstable for now
  Thanks for testing it! Star Citizen telemetry stays on unstable a while longer. The stable version does not get it yet. Its reactions now come in six groups by situation: combat, ship alerts, flight and travel, survival, body needs and wallet. Body needs are off by default. Tell us on Discord how it works for you.

• Less load on your game
  Some of you saw the game stutter while your Wingman listened or looked at the screen. Wingman now takes less from it:

  • Parakeet Ultra replaces the old speech recognition model. It needs a lot less graphics memory, is faster and makes fewer mistakes in all languages, most of all in long sentences and over radio noise.
  • Parakeet frees the CPU between sentences and uses at most 4 CPU threads. On CUDA it takes only the graphics memory it needs.
  • Screenshots for Vision AI are shrunk in half the time. On Windows they run at a lower priority, so the game comes first.
  • New in Vision AI on Windows: Use DXCam capture takes the screenshot through DirectX. Off by default. Try it if your game stutters when your Wingman looks at the screen. If it fails, the standard capture takes over.

• Decision Model: pick who decides
  Jev's switches now have their own section: Settings → General → Decision Model. You pick who answers:

  • Wingman subscription: as before, your plan picks the model.
  • OpenRouter: every decision model OpenRouter offers, with its price. Uses your OpenRouter key.
  • Remote: your own server with TypeSafe's Jev API. It very likely decides worse than Jev in the subscription, which uses hardly any of your usage.

New

1. Skill settings in groups
   Skills with many settings show them under headings you can fold open. Many groups have their own Enabled switch in the heading. Star Citizen Events and Radio Chatter use it, with clearer names.

Improved

• Parakeet on CUDA tells you that it uses graphics memory, and to switch to CPU if you see frame drops.
• On Windows, Settings → General tells you that Windows Game Mode can slow Wingman down while you play, and where to turn it off.

Fixed

• A microphone that is slow to start, for example a headset behind SteelSeries Sonic Studio, was reopened every 3 seconds and sometimes never worked. Wingman now gives it more time with every try.

Found something? Tell us on Discord.

The first big update since the v3 launch: one key now reaches all your Wingmen, Wingman takes less from your game, speech recognition makes fewer mistakes, and you choose who makes the quick decisions. Your HUD goes straight into your stream, and your Wingman can speak up on its own. Plus many fixes.

The biggest news

• One key for all Wingmen
  Hold the period key . and say a Wingman's name in the first 8 words. Without a name your ⭐ Wingman answers, or the Decision Model (Jev) picks the most likely one. Change the key in Settings → Speech-to-text. Each key belongs to one owner: the one you record last wins, and a message tells you where it moved from. Keys that are bound twice right now go back to their defaults with the update.

• Less load on your game
  Some of you saw the game stutter while your Wingman listened or looked at the screen. Wingman now takes less from it:

  • Parakeet Ultra replaces the old speech recognition model. It needs a lot less graphics memory, is faster and makes fewer mistakes in all languages, most of all in long sentences and over radio noise.
  • Parakeet frees the CPU between sentences and uses at most 4 CPU threads. On CUDA it takes only the graphics memory it needs.
  • Screenshots for Vision AI are shrunk in half the time. On Windows they run at a lower priority, so the game comes first.
  • New in Vision AI on Windows: Use DXCam capture takes the screenshot through DirectX. Off by default. Try it if your game stutters when your Wingman looks at the screen. If it fails, the standard capture takes over.

• The HUD as a Browser source in OBS
  Add a Browser source in OBS with the address http://127.0.0.1:7862/obs and put it on top of your game capture. You find the address with a copy button under Settings → HUD Overlay. Only want the chat or the info panel? Use http://127.0.0.1:7862/obs/chat or http://127.0.0.1:7862/obs/info. Thanks to JayMatthew!

New

1. Decision Model: pick who decides
   Jev, the model that makes the quick decisions, now has its own section: Settings → General → Decision Model. Each decision has its own switch: commands, picking the Wingman, misheard names and helping skills choose. You pick who answers:

   • Wingman subscription: as before, your plan picks the model.
   • OpenRouter: every decision model OpenRouter offers, with its price. Uses your OpenRouter key.
   • Remote: your own server with TypeSafe's Jev API. It very likely decides worse than Jev in the subscription, which uses hardly any of your usage.

2. Proactive Comments: your Wingman speaks up on its own
   When it is quiet for a while, your Wingman says something on its own about your conversation or what is happening on your screen. Turn on the Proactive Comments skill for your Wingman and ask it to turn the comments on. Thanks to teddybear082 for the skill and to sneepe, who came up with the idea!

3. ElevenLabs speaks with Eleven v4 Turbo
   The new fast model from ElevenLabs sounds more alive and understands audio tags. The Speed slider makes your Wingman talk slower or faster. The update resets the ElevenLabs model and voice settings to the new defaults. The voice you picked stays.

4. Skill settings in groups
   Skills with many settings show them under headings you can fold open. Many groups have their own Enabled switch in the heading. Star Citizen Events and Radio Chatter use it, with clearer names.

Improved

• Every key you bind can now also be a combination with Ctrl, Shift or Alt, a mouse button or a joystick button.
• Esc and the F keys can now be clicked while recording a key.
• The Avatar Studio asks you to describe your Wingman first and writes the detailed prompt from that.
• The Avatar Studio and the Image Generation skill show how much of your monthly usage one image takes.
• Radio Chatter can wait up to 30 minutes between messages.
• Parakeet on your own server takes a model name and an API key, for servers like oMLX or LM Studio.
• Parakeet on CUDA tells you that it uses graphics memory, and to switch to CPU if you see frame drops.
• On Windows, Settings → General tells you that Windows Game Mode can slow Wingman down while you play, and where to turn it off.

Fixed

• Browse next to path fields opens the folder dialog again.
• A port in the host of the Support Model on your own server no longer stops Core from starting.
• Repeating timers no longer fail with a JSON error.
• When the model calls a skill with wrong arguments, it learns the right ones and tries again right away.
• An unreadable file in the audio library no longer crashes playback.
• Enhance backstory no longer gets stuck.
• Local MCP servers that Core starts with a command (STDIO) start again.
• Vision AI tells you when your model does not understand images, instead of answering empty.
• Pocket TTS no longer crashes when you stop or switch it while it speaks.
• When Wingman Pro refuses a text, you now see the reason.
• A backup folder inside a config folder no longer lists your Wingmen twice.
• An empty answer from the model no longer breaks the turn.
• An expired or already used sign-in link now tells you so instead of failing with an error.
• The microphone test (Try it) in the settings now shows the real level and speech score of what you said.
• Streaming with ElevenLabs now works on plans below Pro too.
• The ElevenLabs settings hardly ever show "too many requests" any more.
• A renamed Wingman shows its new name in the chat and on the HUD right away.
• A local model without tool support, like gemma3:4b on Ollama, answers without skills instead of failing every message.
• After a command, a tool call some models write out as text is no longer read aloud.
• After a command, an answer with a pause, a spelled-out code or a <sigh> stayed silent. It is spoken again.
• Core started from a terminal ends again after Ctrl+C.
• A microphone that is slow to start, for example a headset behind SteelSeries Sonic Studio, was reopened every 3 seconds and sometimes never worked. Wingman now gives it more time with every try.

Found something? Tell us on Discord.

New

1. Star Citizen telemetry
   Star Citizen Events now reacts to your ship and your pilot: shields, hits, locks, quantum jumps, fuel, health and more. Ask "How are my shields?" and your Wingman answers from the real values. Needs Star Citizen 4.10.2. On Windows, quit the RSI Launcher completely (tray too) and start it again once. Switch and status: Settings → Star Citizen. Thanks to Mallachi!

   • Some moments are seen by the game log and the telemetry, like quantum arrivals or the pilot going down. You pick which one speaks, under Game log or telemetry, and your Wingman says it once. While the game sends no telemetry, the game log speaks.
   • No collision warning while your landing gear is down.

2. One key, one owner
   The key you record last wins, and a message tells you where it moved from. Important: keys that are bound twice right now go back to their defaults with this update.

3. Jev: one switch per decision
   Jev now has its own switch for commands, picking the Wingman, misheard names and helping skills choose. Since 3.2.15 you find them under Settings → General → Decision Model.

4. HUD chat and info as their own sources
   http://127.0.0.1:7862/obs/chat and http://127.0.0.1:7862/obs/info show just that panel, sharp at its own size. Thanks to JayMatthew!

Improved

• Avatar Studio: describe your Wingman first, it writes the detailed prompt from that.
• Avatar Studio and Image Generation show how much of your usage one image takes.
• Radio Chatter waits up to 30 minutes between messages.
• Parakeet on your own server takes a model name and an API key (secret parakeet).

Fixed

• Saving a new global, Shut up or Mute key no longer fails, and a new mouse or joystick key works without a restart.
• A renamed Wingman shows its new name right away.
• A local model without tool support, like gemma3:4b, answers without skills instead of failing, and a notice tells you once.
• After a command, tool calls written as text are no longer read aloud.
• After a command, an answer with a pause, a spelled-out code or a <sigh> stayed silent. It is spoken again.
• Core ends again after Ctrl+C.

Found something? Tell us on Discord.

New

1. One Push-to-Talk key for all Wingmen
   Hold the period key . and say a Wingman's name in the first 8 words, for example "Computer, put the landing gear down". Without a name your ⭐ Wingman answers. Without a ⭐ Wingman, Jev picks the most likely one when Fast decisions are on. For that, Jev knows each Wingman's description, skills and commands and who answered last. So even an "And where exactly?" reaches the right Wingman. Also, every key you bind can now be a keyboard key, a mouse button or a joystick button, also with Ctrl, Shift or Alt. That goes for each Wingman's Push-to-Talk, the key for all Wingmen, the Shut up Key and the Mute Key.

2. The HUD as a Browser source in OBS
   Streaming? Add a Browser source in OBS with the address http://127.0.0.1:7862/obs. It shows the HUD on a transparent background at the size of its screen, and you simply put it on top of your game capture. You need no window or desktop capture for it. The address is under Settings → HUD Overlay next to Test Server, with a copy button. With Streamer Mode on, General points you to it as well. When Core restarts, the source reconnects on its own.

3. Proactive Comments: your Wingman speaks up on its own
   When it is quiet for a while, your Wingman says something on its own. It picks up your conversation or, with Look at the screen on, what is happening on your screen right now. Turn on the Proactive Comments skill for your Wingman and ask it to turn the comments on. Or switch on Start automatically. How long it has to be quiet and how often it speaks up in a row is set in the skill. Answer a comment and it still knows what it said. Thanks to teddybear082 for the skill and to sneepe, who came up with the idea!

4. ElevenLabs speaks with Eleven v4 Turbo
   Wingman AI now uses Eleven v4 Turbo, the new fast model from ElevenLabs. It sounds more alive and understands audio tags like [laughs] or [whispers]. The new Speed slider makes your Wingman talk slower or faster, and Wingman AI now tells the model which language you speak. Important: the update resets the ElevenLabs model and voice settings of all your Wingmen to the new defaults, including Stability and Similarity Boost you set yourself. The settings of the old models don't fit v4. The voice you picked stays.

Improved

• On Windows, Wingman AI now tells left and right Ctrl and Alt apart, also in commands you record anew. Your existing keys and commands stay as they are.
• Esc and the F keys couldn't be pressed while recording a key. Now you just click them in the recording dialog.
• The welcome in the channel with all Wingmen now stays in the chat and scrolls up with new messages.

Fixed

• Browse next to path fields, for example Auto Screenshot's default folder, opened no folder dialog. It opens again now.
• With the Support Model on your own server, a host that already contained a port, like http://192.168.1.50:1432, stopped Core from starting. Core now starts anyway, and the log tells you to enter the host without a port. Save the corrected address and Wingman AI connects again right away.
• Repeating timers like "say hello every five seconds" sometimes failed with a JSON error. The Timer skill now accepts the model's arguments in both forms.
• When the model called a skill with wrongly named arguments, for example open folder with a parameter of load folder contents, the call only failed with "error during processing". The model now learns which arguments the skill takes and calls it correctly right away.
• A file in the audio library that was there but could not be read, for example because it was still being copied or contained no sound, crashed playback. You now get a notice, and the play button resets.
• When the model sent the summary of Enhance backstory as a list instead of text, the dialog got stuck. It now arrives as text, and the dialog shows a list as separate lines.
• Local MCP servers that Core starts with a command (STDIO) never came up, the connection failed right away. They start now, and what they report goes to logs/mcp-<name>.log.
• When your Wingman used a model without vision, Vision AI only returned an empty answer. The Wingman now tells you to pick a model with vision.
• Pocket TTS: a built-in voice on a custom model or a download without internet now shows as a hint instead of an error. And if you stop or switch Pocket TTS while it speaks, it waits for the end of the sentence instead of crashing.
• When Wingman Pro refused a text, for example because it was over 5000 characters long, you only saw "Error during TTS playback". You now see the reason.
• A backup folder inside a config folder listed every Wingman twice in the sidebar. Only the top level of the folder counts now.
• When the model sent an answer without content, the turn failed with an error. The Wingman now asks you to try again.
• When the client asked for data before you were signed in or while the config was reloading, Core reported errors. It now simply answers empty.
• Opening a sign-in link from the mail a second time, or only after a while, failed with an error. It now says: "This sign-in link has expired or was already used. Please sign in again."
• The microphone test (Try it) in the settings showed a level of 0 % and a speech score of 0 to 2 %, even when the text came out right. It read the numbers after the test had ended, when you were already quiet. It now shows the level and speech score of what you said.
• With an ElevenLabs plan below Pro, streaming failed with an error, because ElevenLabs only sends its highest audio quality from Pro on. Now Wingman AI streams in the quality your plan gets.
• The ElevenLabs settings often showed "too many requests". Wingman AI now asks ElevenLabs much less often and tries again shortly after when there are too many requests.
• Eleven v4 could not be used with ElevenLabs. Now it speaks.

Found something? Tell us on Discord.

You may not recognise Wingman AI once you move to our latest version. We have been working HARD on the next iteration and left almost no stone unturned. We don’t want to bore you with too many details, so here is the gist of it:

The biggest news

• A brand-new frontend
  Rebuilt from the ground up. Discord-style “channels”, easier Settings, Themes (including Light/Dark Modes) and much more comfortable to use.

• A completely new backend
  Finally allowing us to fix these all-time “no-nos”:

  • Auto login. Yeah, for real! We haven’t had to log in again since we built this. You can now also log in via Discord. Oh, and no more passwords, you’ll see!
  • Auto updates! The v3 setup will be the last one you’ll ever have to download and install manually.
  • We now have access to ALL the models immediately, regardless of the provider. There are many new options to try.
  • Proper subscription and account management: Change (monthly) plans on the fly, cancel anytime, delete your account, and you can do all of this on the new wingman-ai.com website, too.
  • A modern image generation model and a completely overhauled avatar generator and Image Generation Skill for your Wingmen.

• Persistent memory
  Your Wingmen no longer forget everything you talked about the next time you restart them. Their memory is also editable.

• Massive performance improvements
  In all areas: better and faster transcriptions using NVIDIA Parakeet, faster command execution using a decision model (optional, default: off), faster TTS… all with zero setup and preconfigured.

• HUD overlay and Star Citizen Log Reader
  A very cool HUD overlay (by JayMatthew) and Star Citizen Log Reader (by Mallachi), now fully included in Wingman AI Core. You can even use these services in your own Custom Skills. Your Wingmen can now react to in-game events proactively.

• A new and fair usage-based system
  You get a budget based on your plan and then you can decide how to spend it. Do you want to use the expensive providers and models? Go ahead! Do you want to build that fully automated and infinite Twitch/Discord/Wingman loop that keeps your Wingmen talking all the time? This is now allowed as well! We also have a new “support model” that handles easy tasks with cheaper models to save tokens, and Wingman AI gives you tips on how to save tokens by using local providers etc.

  • There is now a real “Free” plan with access to almost all features in Wingman AI. Create as many channels and Wingmen as you like. Use all the providers. Try Skills and MCP servers and then decide whether to subscribe or not. Just like any other plan, the Free plan’s quota resets every month, and we even give you hints on how to save tokens and keep using Wingman AI for free.
  • Pro plan users get 10x the usage, Voice Activation and more voices.
  • Ultra plan users get 30x the usage, Voice Activation and even more voices on top. We do not expect any reasonable Ultra user to ever hit the quota limit, but we’re watching the data closely and will adapt our quotas if needed. Ultra should still feel unlimited unless something is really broken.

• Local PocketTTS now supports multiple languages
  And not just English. Language selection was improved and simplified significantly, so Wingman AI will no longer detect or answer in the wrong language. Local voice cloning is also supported: All you have to do is put a sample audio file in a folder and you’re good to go. PocketTTS is local, so using it doesn’t cost you any tokens.

• Tell a friend
  A new program, replacing our old 7-day trial and discount codes. You get a unique ID/URL you can share with your friends, and as soon as they subscribe to any plan, you both get a usage reset token that never expires and that you can use on any plan.

• v3 Skill Facade
  For developers and vibe coders: excellent agent instructions, allowing you to use your coding agent to create Custom Skills for you. v2 Custom Skills need a migration. Check out our Core repository and tell your AI agent you want to do that, and it will know what to do.

• Windows, macOS and Linux
  Wingman AI is now available on Windows, macOS (well tested) and Linux (very experimental).

And a ton more

That’s just the biggest news, but there are A TON of other new features and fixes, like:

• StarHead MCP completely reworked and great again
• MCP servers now support OAuth
• Commands can now call Skill functions/tools
• Skills and MCP tools can now be disabled independently
• Full changelogs and feature requests are now integrated in the client and no longer need a separate account
• Translations are much better now
• Way better error logging and reporting, allowing us to help with your support cases more efficiently
• TTS now handles numbers and long outputs way better

Try it now

Now is the best time to try out Wingman AI, or to revisit it if you liked the idea before but couldn’t get it to work properly or were overwhelmed by its options.

Found something? Let us know on Discord.

New

1. Your Wingman keeps track of missions and plans
   Take on a mission, agree to a job or set up a meeting, and your Wingman now remembers it on its own, without being asked. When it is done, it closes it and keeps the outcome, for example "Delivered the cargo to Area18, got 40k aUEC". It knows open things in every session until they are done. Where you are right now, your mood or a jump it carries out right away do not go into memory. You no longer need "remember new missions" in the backstory for this. In the Memory tab you filter by Open and Done.

2. One voice picker for every provider
   Edge voices are now in the tile picker too, with filters for language, gender and Multilingual. The list comes live from Microsoft and has readable names like "Ava Multilingual". The star marks favorites, they come first. The eye hides voices, they go to Hidden at the bottom. Favorites and hidden voices apply to this computer only.

3. New default voices for Computer and ATC
   Computer speaks with a female voice in every language, ATC with a male one. Computer now has the AI sound effect. If you never changed a Wingman's voice, it switches along. Restore defaults in the TTS tab resets only voices and sound effects, everything else stays. On the Free plan, Azure now has Xiaoyu Multilingual and Lewis Multilingual instead of Jenny Multilingual, Katja, Denise and Elvira. On the Pro plan, Inworld has Deborah, Sabine and Brian instead of Ashley and Johanna.

4. Terms and privacy policy in the app
   The consent dialog shows both texts right in the app. You can accept once you have scrolled both to the end. In the About window you find the terms, the privacy policy and the cookie policy as their own pages.

Improved

• Radio chatter and game events no longer cut off answers. Answers to you come first, then game events, then radio chatter. An event waits up to 60 seconds until you and your Wingman are done. Radio chatter never talks over anything, and a line that doesn't fit right now is dropped.
• Hotwords only when you say so. Your Wingman no longer adds a spelling on its own, also not for a name that looks misheard. Before, it could rewrite your real name into the name from its backstory, in every configuration. Each new entry shows in the chat, with a note that it sits in the hotwords in Settings.
• Previews sound like the real answer. Every voice preview now uses your pronunciation rules. In the Pronunciation editor the written side plays the word as written, so you can compare. The preview sentences in German, French, Spanish, Italian, Portuguese and Dutch no longer use "Wingman" and "online".
• German voices say "online" right. A new pronunciation list Wingman is switched on once and has rules for "online" and "Wingman".
• Vision AI reads wide screens. Screenshots go out as JPEG. On a 32:9 screen small text stays readable, before it shrank to 3.5 pixels.
• Pocket TTS 4.0.1. On first start Pocket TTS downloads the built-in voices once more and makes your cloned voices once more.
• Commands tab. The header with the view toggle, Add Group and Add Command stays visible while you scroll.

Fixed

• Edge TTS stayed silent because Microsoft changed its service.
• Vision AI failed on busy screens, and the Wingman then said the screenshot had failed.
• After switching configurations, parts of the old one kept running: keys, voice input and MCP servers. When signed out, the old configuration even kept answering.
• Timer reminders never came.
• After Voice Changer or radio chatter, Pocket TTS streaming was off, and every line started after 6 to 28 seconds of silence.
• With Edge, the preview's stop button flipped back at once. On macOS, a preview clicked right after another one broke off.
• New keybinds on a command showed in the list view only after a restart.
• On Linux, sign-in did not finish when Gear Lever or AppImageLauncher had put the AppImage into the menu.
• Truck Simulator 2 Telemetry no longer crashes when mods set single values to zero.

Known issues

• On Retina displays the HUD text is a little soft.
• On Linux the HUD looks smaller on a scaled display, and a native Wayland game in fullscreen can cover it.
• A few items have no in-game name yet and show their technical name.

Found something? Tell us on Discord.

New

1. The new HUD with HUD profiles
   All Wingmen now share one chat and one info panel. In Settings > HUD Overlay you set with HUD profiles where they sit and how they look, per channel. Per Wingman, General has Show messages on the HUD and Can pin to the info panel on the HUD. With Preview and position the HUD you drag them into place right on your screen. The Chat history shows the conversation like a messenger, with pictures and your own lines. The HUD skill is gone for this, and the update moves your HUD settings into a profile.

2. The HUD runs on macOS and Linux
   On Linux under X11, and in Wayland sessions through XWayland. Mic Status runs there too now.

3. Control Windows and Auto Screenshot on macOS and Linux
   Your Wingman now opens, switches and closes apps and windows there too. On macOS it needs Accessibility in System Settings for that. Under Wayland, Vision AI now takes real screenshots instead of black pictures.

4. Feature requests in the app
   With Request feature in the header bar you vote for what we build next, suggest something new and comment. Just paste pictures in. You don't need Canny for this any more.

5. Changelog in the account menu
   At Info > Changelog you read the release notes of every version and page back with the arrows. About is there too, with the version, updates and the community links.

6. Star Citizen Events: one switch per event
   Under the seven groups you now switch each event on its own. New reports: leaving a station and quantum arrivals with any ship. A purchase comes once, when the game confirms it, with its real name and price, for example "7CA 'Nargun', 1226 aUEC". The list of names updates without a new release.

7. UEXCorp: multi-stop trade routes with guidance on the HUD
   A trade route can now have several stops, with purchases along the way and round trips. You say how many stops and system jumps it may have at most. For the best routes you get an exact cargo plan, with notes when money, stock, demand or container sizes limit a purchase. Pick a route by its number and your Wingman guides you stop by stop, while a HUD panel shows the stops. Tell it what you really bought or sold and it keeps calculating with that. At Trade route ranking you choose whether profit per stop, per distance or in total counts. The data comes from our server, loads in a few seconds and stays usable when the server is briefly unreachable.

8. Token count shows the context
   The number next to the Wingman is now the size of the context, not the sum of every request. When an answer needs several requests, the count shows next to it. The tooltip lists each request.

Improved

• Tidier settings. The order is now General, Star Citizen, HUD Overlay, then audio, transcription, TTS and secrets. The Shut up Key is under TTS. On a Wingman the Personality panel is now General, and the backstory tab is Personality.
• Commands only when you want them. Say "next stop is Lorville" or "I fitted new guns" and your Wingman presses no key. Answers are shorter, with the answer first. The update adds the new rule to your own system prompt and keeps your edits.
• Star Citizen Events sound more varied. Each reaction knows the last five and picks a different angle. No more same joke on every event, also with backstories that have their own ALWAYS rules.
• Star Citizen Events wait for you. While you hold push-to-talk, speak or an answer runs, your Wingman does not cut in. Zones, quantum jumps and the hangar come at most every 5 minutes, adjustable from 2 to 20 at Minutes between routine lines. Arrivals and departures always come. The reports no longer show up again in later answers.
• Speech recognition no longer turns normal words into names. "Radar" became "Yadar". Common words now stay as they are. Misheard names like "lorvil" are still fixed.
• Talking over your Wingman. Only runs of three words or more count as its echo. Before, single words went missing from your sentence.
• Filler line. The short line while a tool runs is skipped when it would be ready only after the answer.
• Subscription answers stay in the prompt cache. Each Wingman stays with the same provider. Long conversations get faster.
• Memory model. It downloads from our server, a broken download continues where it stopped, and it happens in the background. You can start right away.
• File Manager is safer. File names stay in their folder, key files are not read, scripts go only to the default folder, and autostart and Wingman's own folders are protected.
• Error reports. They now carry the log of the current session. The text at Send error reports says so. Errors from your own provider and Windows noise are no longer sent.
• HUD shows answers without speech markup like <break>.

Fixed

• On Windows, Core sometimes stopped answering in the middle of a session while it still ran.
• When two active skills had a tool with the same name, some providers rejected the request.
• An instant command ran twice when a phrase was listed twice. The landing gear went up and straight back down.
• Memory stored nothing on fresh installs but said "Memory stored".
• The HUD reconnected twice every turn.
• Push-to-talk was released twice and said "Recording stopped" twice.
• Memory checkpoints got lost and unloading a Wingman failed.
• A Pocket TTS voice without its file now says clearly that it is missing.
• The terminal grew until storage was full. It now keeps the last 500 messages.
• The instant activation switch crashed on older commands.
• Plus many smaller crashes from your error reports.

Known issues

• On Retina displays the HUD text is a little soft.
• On Linux the HUD looks smaller on a scaled display, and a native Wayland game in fullscreen can cover it.
• "Set a course for Hurston" now mostly switches to navigation mode and waits for your jump order instead of jumping right away.
• A few items have no in-game name yet and show their technical name.

Found something? Tell us on Discord.

New

1. Star Citizen in one step
   After the first start Wingman asks: Do you play Star Citizen? If you are already set up, you see the question once after the update. App language, spoken language and Star Citizen are on one screen. One switch sets up everything at once: the Star Citizen Log Reader Service, the pronunciation and word lists, and the matching skills and MCP servers on ATC and Computer. Under Show Skills & MCPs you switch single ones on or off.

   There is now one game folder for every feature and skill. Wingman finds it by itself, in the RSI Launcher's log and in the usual folders. Wingman only reads files that Star Citizen writes on your computer. It does not read the game's memory, does not change the game and does not upload any files.

   Later you find all of it under Settings > Star Citizen. The log reader moved there from General.

2. Jev is off
   Fast decisions (Jev) is switched off for everyone with this update, and new installs start without it. Many of you did not know that Jev decides things. When Jev got it wrong, it looked like a broken microphone or a Wingman that does not listen. If you want Jev, switch it back on under Settings > Support Model. Your setting for commands stays as it was. The switch now also says what Jev decides.

3. No ⭐ Wingman out of the box
   New installs no longer have a ⭐ Wingman. With voice activation you say a Wingman's name, and that Wingman answers. If one should answer without its name, give it the ⭐. If you already have a ⭐ Wingman, you keep it. The tooltip on the ⭐ now says what a click does.

Improved

• The What's new window also shows older versions: page back with the arrows at the top right. From the next update on, the update notice has a Show changes button. You read what is coming before you install, across several versions too.
• Fields for folders and files have a button that opens your system's picker, for example in ATS/ETS Telemetry, Auto Screenshot and File Manager.
• On new installs the Computer comes with Star Citizen Events and web search. The Star Citizen question switches Galactapedia on for the Computer only, ATC sticks to traffic control. The Accountant stays off until you switch it on.
• Jev's check for misheard names waits 1.5 seconds at most. When it fails, it pauses for a minute. "kinda", "sorta", "gonna", "wanna" and "gotta" no longer turn into names like "Kinga".
• Updating from 2.1.1: the shipped Wingmen get the skills they have been missing since then, and four commands with typos ("Missle") get the right text. Your own Wingmen stay as they are. Afterwards Wingman removes the folders of the versions it skipped.
• The chat jumps straight to its end when you come back from the Wingman editor or switch Wingman, instead of scrolling through the whole conversation.
• macOS: Wingman now needs macOS 14 or newer. On macOS 11 to 13 the app installed, but then Core crashed. Now macOS says so before the app starts. The app itself only has the microphone permission.

Fixed

• A Wingman you switch off stops listening and speaking right away. Before, it kept answering until the next restart. Changes to a switched-off Wingman are saved, and after switching it on it has its memory again.
• With Jev, voice activation dropped some sentences meant for your Wingman, for example "we got the Endurance blueprint". Now your Wingman answers every sentence, and the answer starts about 0.8 seconds earlier.
• Fresh German install on Windows: the first answer stayed silent, a dialog asked for the secret "wingman_pro", and only a restart helped. Now your Wingman speaks on the first answer.
• Changes to volume, provider and model in the chat input were lost on restart.
• The ⭐ could not be set again after it was removed.
• Quitting on macOS and Linux left a llama-server of about 300 MB running every time.

Known issues

• German Pocket TTS voices still say some English and French words the German way, e.g. "Marge". You can fix single words under Settings > Pronunciation.

Found something? Tell us on Discord.

New

1. Long answers with Pocket TTS sound right
   Long answers no longer repeat the last word of a sentence, squeak at the end or start with a hiss. The pauses between sentences are short and even. German voices keep one voice from the first to the last sentence instead of sounding like a new speaker every few lines.

2. Retrained voice models
   German, French, Spanish, Italian, Portuguese and Dutch use Kyutai's retrained models. German makes about half as many mistakes as before. English Standard uses a newer model that no longer repeats the end of an answer ("an hour an hour").

   On the first start Wingman downloads the new models and makes every Pocket TTS voice once more for them. The first answer of each voice takes a few seconds longer. To do it for all voices at once, use Precompute all custom voices in the Pocket TTS settings.

3. Numbers the way people say them
   Large amounts are rounded for the voice: "689,482,137 aUEC" is spoken as "about 690 million aUEC". The chat keeps the exact number. German numbers no longer get stuck in a loop. Spanish, Italian and Portuguese say menos and meno for negative numbers.

4. Lists read like sentences
   When list items are full sentences, your Wingman reads them as sentences with a short pause in between, without an "and" squeezed in before the last one. Short lists like "food, water and fuel" stay as they are.

Improved

• The Star Citizen pronunciation list now has French, Spanish, Italian, Portuguese and Dutch lines for aUEC, ATC, places and ships.
• Inworld with your own key: the voice picker shows gender and descriptions, like with the subscription.
• The voice picker marks Pocket TTS voices whose recording is gone. Outdated voices still speak with their old copy. Unavailable voices cannot be picked. Put the recording back into the Custom Voices folder and Wingman makes the voice again.
• All four languages of the app are reworded: one term for each Wingman concept, and German, French and Spanish always use the informal you.
• The sign-in screen welcomes you and says what happens to your account.

Fixed

• A crack about three seconds into answers on audio interfaces that are microphone and speakers in one device, such as the Audient EVO4, on macOS.
• Voice Changer and Radio Chatter no longer fail on every turn when a Pocket TTS voice in their list was deleted. They skip it and tell you which Wingman's list to clean up.
• Inworld with your own key: the language filter of the voice list works again.
• Precompute all custom voices works on the first click after you switch the Pocket TTS model.

Known issues

• German Pocket TTS voices still say some English and French words the German way, e.g. "Marge". You can fix single words under Settings > Pronunciation.

Found something? Tell us on Discord.

New

1. Azure voices are back, Inworld stays
   The Wingman Subscription speaks with two providers again. Azure brings hundreds of voices in every language, Inworld the most expressive ones. You switch per Wingman under Text to Speech. Azure voices use about 3x as much of your allowance as Inworld, the tiles tell you which is which.

2. A new voice picker
   Picking a voice is a gallery now, for Azure, Inworld and Pocket TTS alike: search by name or description, filter by language, gender and multilingual, and listen to any voice before you choose it. Voices your plan does not include are shown but cannot be picked, so what you set is always what your Wingman speaks.

3. Tell a friend
   You have a personal code and link on your account page. When a friend subscribes with your code, you both get a Usage Reset. New subscribers enter the code right before checkout, and the account page shows how many friends signed up and subscribed.

   Not ready for testing yet. Tell a friend needs the new website: the invite link does not lead anywhere until then, and the reward only arrives 14 days after the friend's first payment. Please leave it alone for now, we will tell you when it is live.

4. Usage Resets
   A Usage Reset starts a fresh month of usage right away, whenever you want it. You earn them through Tell a friend, they stack, and they never expire. Your account page shows how many you have and a button to use one; it asks before it starts the new month.

5. Plans in one place
   Your subscription, the plans and the full comparison live in one dialog, reachable from every upgrade hint and from your account page. Switching plans, pausing and cancelling happen there too.

6. German, French and Spanish voices by default
   Pick your spoken language and your Wingmen speak it with native voices, in Pocket TTS, Azure and Inworld. Players who speak a language other than English now start with Azure voices instead of Pocket TTS.

Changed

• Fewer locks. Every plan can create Wingmen, use commands, skills, MCP servers and the avatar studio. Only Voice Activation and the Inworld voices still need Pro or Ultra.
• Usage is simpler. Free has 1x usage, Pro 10x, Ultra 30x. Every plan includes two Azure voices per language; Pro adds all Azure voices and two Inworld voices per language; Ultra includes every voice.
• You never lose what you paid for. When a plan ends or you switch down, your budget and voices stay until your current usage month is over.
• The usage meter shows when your month starts over, with the time of day.
• One usage badge everywhere: a lightning bolt with 1x, 2x or 3x, green to red.

Improved

• Elite Dangerous and the Star Citizen Galactapedia are available as MCP servers for every subscriber, already configured.
• The account page shows your usage and your subscription together, with the plan dialog one click away.
• Wingmen whose configured voice is not in your plan move to the closest free voice of the same language and gender when you sign in, and tell you so.
• You can delete your account on the account page. It cancels your subscription, removes your data and signs you out. Your usage stays in our statistics without any link to you.

Fixed

• Speech over the subscription now reports a used-up allowance as such instead of a playback error.
• The account menu no longer hides behind the page toolbar.
• Long language lists in dropdowns scroll instead of running off the window.

Known issues

• The plan dialog still shows the old pricing table and the old plan descriptions from our website. They do not match the new plans yet; the website update is on its way.

Found something? Tell us on Discord.

New

1. Your Wingman follows your Star Citizen session
   Wingman now reads Star Citizen's Game.log by itself. The new Star Citizen Events skill uses it: your Wingman reacts in character to missions, zones, injuries and trades, in your language, in the chat, on the HUD and out loud. It also knows where you are and which missions you have. You choose which kinds of events it talks about and how often. Switch the reader on or off and set your game folder under Settings > General > Star Citizen Log Reader.

2. Star Citizen Accountant now ships with Wingman
   The Accountant books your trades and rewards straight from the log, for all your Wingmen in one set of books. A dashboard shows your money, cargo, operations and fleet on your computer, or on your phone via QR code.
   Did you install the SC Log Reader or the Accountant as custom skills? Wingman replaces them with the bundled versions and keeps your existing books.
   Both skills are by Mallachi, who also keeps the log rules up to date on GitHub. Thank you!

3. Avatar studio: you see and edit the prompt
   Your Wingman writes the image prompt first. You can change every word, and the image model gets exactly that text. One switch picks between a new image and changing the current one. To change one, you only describe what should be different.

4. Updates in one click
   When a new version is out, a notification offers to install it, shows the download progress and restarts Wingman AI into the new version.

Improved

• Your Wingman forgets less. Tool results like prices and stock now stay complete for at least four turns, and with Auto-summarize conversations off nothing is removed at all. All limits follow the context window of your AI model. The old settings for tool compacting are gone, Auto-summarize conversations is the one switch left.
• A stop button appears in the header whenever a Wingman speaks, on every page. While push-to-talk records, a "Listening" pill and a red ring around the avatar show it.
• The usage meter moved into the profile menu and shows the day your allowance starts over.
• HUD: your Wingman can bring back a panel the way it was before it was changed or removed.
• Skills that need the HUD or the Star Citizen Log Reader stay off until that is on, with a hint that leads you there.
• We went through all bundled skills and fixed bugs in 18 of them. For example, Audio Device Changer no longer resets your microphone on every answer, and MSFS had "on the ground" and "in the air" swapped.
• For skill authors: skills can open a dialog with text and an image (self.wingman.ui.show_dialog) and read the Star Citizen log via self.wingman.sc_gamelog. The skill docs are updated.
• Windows: audio devices show their full name instead of the first 31 characters.

Fixed

• Linux: mouse actions like "Initiate Quantum Jump" no longer need root.
• With your own OpenRouter key, Wingman sometimes thought a model could not use tools, so skills and commands did not work with it.
• Local AI: when another app uses the port, Wingman starts its local AI servers on a free one.
• Wingman AI now shuts down cleanly before an update restarts it.

Found something? Tell us on Discord.

New

1. One spoken language for everything
   Pick your spoken language once in Settings > General. Speech recognition, your Wingman's answers and its voice all follow it. On first start, Wingman now asks for your app language and your spoken language separately. Changing the language also reloads the appropriate models and warms them up.
   Custom skills can now read the Wingman's language via self.wingman.language in the skill facade and incorporate it into their own prompts accordingly. For core skills, this has already been implemented.

2. System 1/Jev: AI commands in half a second
   A small, fast decision model now checks every request for a command before your main model even gets it. AI commands fire after about 500ms on average, and the spoken confirmation comes about a second sooner. The catch: every reply that is not a command now takes about 500ms longer. Jev can also enable Commands for models that are usually bad at calling them.
   If you use Wingman mostly as a VoiceAttack replacement, leave it on. If you mostly talk or roleplay with your Wingman, turn off Decide commands before asking the main model in the settings. There are other System 1 helpers, like better name recognition and these always stay on automatically.

3. Command descriptions
   Every command has a new description field right under its name. A short sentence like "Raises the landing gear" helps Wingman pick the right one. The shipped Wingmen already have AI-generated descriptions which you are invited to make better. These descriptions help both Jev and the main model (if Jev is turned off).

4. Pocket TTS 3.3.0: much better voices beyond English
   New models for German, Spanish, Italian, Dutch and Portuguese, a faster French model, and a new English one. Non-English voices now make far fewer mistakes: in our German test, misspoken words dropped from about 11% to 6%, almost as good as English. On top of that come 145 new voices recorded by native speakers. You pick them in a new dialog with search, a gender filter, a short description of each voice and a preview button.

5. Pronunciation rules aka "reverse hotwords"
   Your Wingman keeps saying "aUEC" wrong? Under Settings > TTS > Pronunciation you can tell it how to say a word. For Star Citizen there is a ready-made list of game terms. Please feel free to extend and PR that.

6. StarHead is great again
   StarHead MCP got a huge upgrade from Timo and delivers much more and better data now. Please try it out! This also works in Wingman 2.1.1 and any other version.

7. Filler words are back
   When a tool takes a moment, your Wingman says a short line in your language, like "On it!", so you aren't left in silence. You can turn it off in Settings > General.

Improved

• Voice previews with SFX now speak your spoken language, and voice pickers warn you when a voice doesn't speak it.
• If you are curious, each Wingman message can show how many tokens it used, and the status bar shows the total. You turn this on in Settings > General.
• Linux: The AppImage shrank from about 4.3 GB to 2 GB. It now finds your PipeWire and Pulse audio devices (USB headsets included), and hotkeys work under Wayland.
• MacOS: Wingman now asks for the microphone, accessibility and screen recording in a clear onboarding step. You can check them later in the settings.
• Pocket TTS sounds cleaner: no more clicks between sentences and no cut-off last syllable. The first sentence comes as fast as the rest.

Fixed

• Streamed speech sometimes skipped a piece of a sentence.
• Volume and radio effects were applied twice.
• Wingman sometimes "corrected" ordinary words into game names you never said.
• Backspace in the chat input no longer jumps back a page, and you can type while voice activation is recording.
• Since June, Parakeet quietly ran on the CPU on Windows and Linux, even with an NVIDIA card. It now uses your GPU again, on CUDA 13. You need NVIDIA driver 580 or newer and a GTX 16xx, RTX 20xx or newer card. Older setups keep running on the CPU.

Found something? Tell us on Discord. Uncaught exceptions should now reach us automatically, unless you turn them off under Settings > Send error reports.

This entry sums up everything that shipped to unstable testers since 2.1.1, newest first. Polished release notes will follow with the official release.

Updates, done right (3.1.5)

• Fixed a Windows packaging issue that could crash Wingman AI on startup after updating.
• The installer now cleans up leftovers from previous versions.
• Updates no longer install silently: a red dot shows up when one is available and the new Settings → About page lets you install it when you are ready — with this changelog right inside the app.

Sturdier configs & voice activation (3.1.4)

• Complete rework of config tracking — no more duplicated or "resurrected" configs after updates.
• Improved voice activation with a new status skill.
• Many crash fixes and UI polish across the board.

Skills 3.0 (3.1.3)

• Skills now declare compatibility — outdated skills are disabled safely instead of breaking your Wingman.
• Skill functions as command actions: trigger skill features directly from your custom commands, including instant responses without an AI roundtrip.
• HUD overlay: images, background opacity, and free positioning.

A fresh look (3.1.3)

• 12 curated themes with light and dark mode — pick yours in Settings → General.
• Redesigned provider selection, first-launch setup, and pricing page.

PocketTTS 2.0 (3.1.3)

• New model picker, per-voice preloading, and preview phrases.
• Smoother voice-cloning workflow.

Better Local AI (3.1.0 – 3.1.3)

• Big quality pass on the local support model: smarter defaults and heavily tuned settings for Persistent Memory extraction and summaries.
• New Local AI Lab: chat with, benchmark, and tune your local models right inside the app.

Faster local speech recognition (3.1.0)

• Parakeet is the new default speech recognition — faster and more accurate than Whisper for most voices, running 100% locally.
• Switch between local engines (Parakeet / FasterWhisper) with a click; models download automatically in the background.

Wingman AI goes multi-platform (3.1.0)

• Native builds for macOS (Apple Silicon) and Linux, alongside Windows.
• macOS guides you through the Accessibility permission it needs for your push-to-talk keys.

Major Features

PocketTTS: 100% Local Text-to-Speech with Voice Cloning

Did we mention PocketTTS also supports instant voice cloning?
Just place a clean 30-second .wav file (or even faster: a .safetensors sample) into the new custom_voices directory, and boom — you’ve cloned that voice and can use it in Wingman AI immediately.
We’re also bundling several voices with open licenses. Additionally, the community can share custom voice samples for cloning in our local-models channel.

If you have a dedicated “AI Server” in your network and want to offload PocketTTS from your gaming rig, check out the OpenAI-compatible PocketTTS Server, maintained by @teddybear082 and @shackless. It can be connected to Wingman AI using our “OpenAI-compatible” provider.

The only downside for now is that PocketTTS currently supports English only. More languages are planned, and as soon as they’re available, PocketTTS will become our new default provider.
PocketTTS is implemented as a first-level provider in the client, so you can simply click on the TTS provider in your Wingmen settings and select it — that’s all.

HUD Skill

@jaymatthew has been working on this one for a while, and now it’s ready to become an official Core Skill.
The HUD Skill renders various in-game information from Wingman AI and is highly customizable.
Go to the global Wingman AI Settings view and enable the “HUD Overlay” checkbox. Then go to your Wingman and add the new “HUD Overlay” Skill for each Wingman that should push data to the HUD.
Give it a try!

Better Commands UI with Groupings

A highly requested community feature is finally here: You can now create groups for Commands in the client.
We also rebuilt the frontend and added several convenience updates:

• New list layout showing which Commands have instant activation, response phrases, or are AI-executed.
• Added a new grid view toggle showing even more information than the list view.
• Commands can now be multi-selected and deleted, reordered or regrouped in batches
• Command details now open in a drawer view (like the Audio Library) instead of a new page, so you won’t lose your scroll position when editing.
• NumPad bindings are now displayed and detected/simulated properly.
• @xul completely overhauled our example bindings for ATC and Computer. The new defaults include more Commands, all properly named and grouped — just as you’d expect when playing Star Citizen.
• Commands now have a new field called “Additional Context” which is sent to the LLM after a command has been executed. You can use it to give it some static information and guide its reply.

We can’t auto-migrate Commands if you have custom bindings.
So if you’re upgrading from an older Wingman AI version and want the new default bindings and Commands, hit the new “Restore” button on both ATC and Computer to reset them to the new defaults.

More Features

• Added a new “Restore” button for the default Wingman, allowing you to reset to default settings if you’ve made changes or encountered issues.
• Added a new “Uninstall” button for Custom Skills. Use this before installing a new version of a Custom Skill. It removes all files and detaches the Skill from all your Wingmen, which is important if the new version introduces breaking changes. Then simply download and extract the updated Skill version to your custom_skills directory in %APPDATA% and attach it to your Wingman again in the client.
• 11Labs v3 model is now supported.
• Added new custom property type “color” for Skill developers which is rendered as a color picker and returns a HEX string in the client
• Migrated the Windows Setup Wizard to Inno Installer due to size limits. It’s now an .exeinstead of a .msi and looks a bit different but don’t worry. Just install 2.1.1 over your 2.0.0 instance as usual.
• Our battle-tested gpt-4o-mini model has been deprecated and removed from Azure. Therefore and because we think it’s the best model available for Wingman AI right now, we are moving on to the newer gpt-4.1-mini. Both Pro and Ultra users get it automatically, effective immediately. It’s much more expensive for us but also a nice upgrade for both plans, so we are not sacrificing quality here.

Bugfixes

• Fixed an OpenRouter validation error caused by their API change.
• Prevented deletion of system Commands (ResetConversationHistory).
• Improved UI rendering and TTS handling of Markdown tables and lists.
• Improved thread handling and performance, especially when listening to joystick input
• Removed broken drag&drop for Command actions in favour of a more stable and consistent up/down arrow solution to reorder. Also all Command actions are reorder-able now
• Fixed custom properties sometimes “jumping” and switching positions in the client
• Fixed QuickCommand Skill adding duplicate instant activation phrases leading to duplicate commands executions
• Fixed hallucinated tool calls and reasoning for *-latest Google Gemini models
• “Enhance backstory” can now process larger prompts
• Audio Library now correctly detects when a preview playback ends
• MSFS Skill now matches command names intelligently
• Improved Markdown rendering in lists and tables
• Fixed Inworld voice selection in Skill’s custom properties
• Starhead and Datetime MCPs can now always be toggle on/off, regardless of your subscription plan
• Fixed Voice Selection controls losing their value if the same value as before was selected
• Fixed continuation bug in the “Reset Password” workflow
• Config validation errors are now logged less verbose and easier to read
• Potentially fixed infamous “white screen” issue that has been haunting us for years (?)

Major Features

MCP (Model Context Protocol) Support - The Future is Here!

Wingman Ultra now supports MCP servers, which means you can connect external tools and services to your Wingmen. Think of it as a plugin system that works with any MCP-compatible tool. The really cool thing is that MCP servers can either run locally on your machine and interact with programs you might already be using (like Notion, Figma etc.) or they can be hosted remotely (if they’re stateless and don’t need local file access). It’s basically like a Skill but on a different server, meaning it can be updated independently from Wingman AI and it’s super easy to connect to.
MCP is like the official protocol for Skills that didn’t exist yet when we introduced Skills to Wingman AI.

What does this mean for you?

• Access to more data: Connect to web search, game databases, and other programs or services
• Community-driven: Developers can create MCP servers in any language, and you can use them in Wingman AI. There are already thousands of MCP servers out there!

Built-in remote MCP servers, hosted by us for you:

• Date & Time Utilities: Current date time retrieval and time zone conversions (free for everyone)
• StarHead: Star Citizen game data (ships, locations, mechanics) (included in Ultra but still enabled for everyone in our Computer example Wingman)
• Web Search & Content Extraction: Brave and Tavily search (included in Ultra)
• No Man's Sky Game Data: Game information and wiki data (included in Ultra)
• Perplexity AI Search: AI-powered search (bring your own API key)
• VerseTime: This one is new and was created by popular demand from @ManteMcFly. It’s basically like “Date & Time Utilities” but for Star Citizen locations so that you can plan your daylight/night time routes more easily.

We’re planning to add more MCP servers in the future and who knows… maybe we’ll have a “discovery store” at some point allowing you to discover new MCP servers more easily.

Progressive Tool Disclosure - Smarter and Faster

We've completely reworked how your Wingmen discover and activate Skills and MCP servers. Their descriptions and meta data no longer pollute the conversation context as soon as they are enabled. Instead, the Wingmen get a very slick discovery tool allowing them to find capabilities (which are either Skills or MCP servers - the LLM doesn’t care) ad-hoc. Only when discovered and activated, Skills and MCPs then add information to the context - saving tokens and precious response time. We tested this a lot and it works way better, especially with smaller (local) models.

Inworld TTS (experimental)

You wanted more voices and now yet get them!

• Included in Wingman Ultra (for now!)
• High-quality voices with inline emotional expressions (anger, joy, sadness, coughing etc.)
• Very low latency around 200ms, making it the fastest provider today
• Way more affordable than ElevenLabs
• If you bring your own API key (BYOK), you can also use voice cloning

Word of truth: While Inworld is 100% our new provider of choice for English TTS output, it unfortunately isn’t that great in other languages yet, especially in German. They only have 2 voices so far and these aren’t amazing. So if your Wingman responds in German, ignore it for now!
We have to monitor the usage and see if it “fits” into our Ultra plan. If it doesn’t work out, we might have to remove it later or add a new pricing tier for it - just so you know.

CUDA bundling and auto-detection

CUDA is finally bundled with our release and FasterWhisper now automatically detects and uses it if you have a compatible NVIDIA GPU. This is huge because we know many of you struggled with this or just didn’t bother (making everything slower than it needed to be) - now it’s solved! No need to install outdated drivers or extract and copy .dll files anymore.

UI improvements

We’ll never stop improving our client and apart from various minor improvements, this time we have completely redesigned all of our AI and TTS provider selectors. We now show you what the providers actually do and also which capabilities their models have. We also added search, filters and fancy recommendation badges to the mix, so picking the best provider and model has never been never easier.

More Features

• Improved AI prompts: We’ve completely overhauled our system prompt and also the backstories of our Wingmen. There is also a cool new button in our UI that turns your current backstory prompt - no matter how good or bad it is - into our new recommended format and makes it use the latest prompt engineering techniques. You can now literally start a backstory with one sentence and turn it into a magic gigaprompt with a single click of a button. We highly recommend to use it for your migrated Wingman - it will make them better!
• New models and better defaults: We’ve added gpt-5-mini for our Ultra subscribers and updated many default models across providers (e.g. gemini-flash-latest for Google). We’ve also added xAI (Grok) as BYOK provider and all of their models. In turn, we removed the obsolete mistral and llama models and also gpt-4o (outdated, slow and expensive) from our subscription plans.
• Groq (not xAI/Grok with a “k”) can now be used as STT (transcription) provider which is really fast (400-500ms) and cheap for a cloud service. We recommend to test it, especially if you don’t have a NVIDIA RTX GPU or are running Wingman AI on a slow system you don’t want to run local STT on.
• Mistral models are now loaded dynamically, so you always get the latest available models
• Wingmen can now be cloned into any context in the UI
• You can now bind a joystick key as “shut up key” (cancelling the current TTS output)
• Enhanced TTS prompts: Added tts_prompt parameter for OpenAI-compatible TTS providers like coquiTTS, allowing them to use Inworld-style inline emotions (if supported)
• Audio library moved: The sound library is no longer in the versioned working directory, meaning it can easily be migrated and you no longer have to copy it manually. If you had sounds in 1.8, they’ll just be available in 2.0.
• Documentation: Added detailed developer guide for creating custom Skills (see skills/README.md) and added AI-optimized guidelines for AI agents working in the codebase (.github/copilot-instructions.md)

Bugfixes

• Fixed Wingmen thinking they are called "Wingman" unless stated otherwise
• Fixed Voice Activation vs. remote whispercpp conflicts
• Fixed streaming TTS with radio sounds for OpenAI-compatible TTS providers
• Fixed migrations not respecting renamed or logically deleted Wingmen (once again)
• Improved instant command detection to exact match (prevents false positives like "on"/"off")
• Fixed various UI glitches and issues with AudioLibrary and previewing sounds
• Local LLMs can now receive an API key if needed
• OpenAI-compatible TTS now receives the output_streaming parameter and can load voices dynamically (if supported).
• Very long Terminal messages now wrap the line correctly and no longer break the UI layout
• “Stop playback” now also shows the “shut up key” binding
• Prevent “request too long” client exception after excessive use
• Wingman name can no longer contain of just whitespaces, breaking everything
• Fixed Wingmen with same name prefix all being highlighted as “active” when configuring

Breaking Changes & Migration

• validate() is now only called when the Skill is activated by the LLM, unless you set auto_activate: true in the default config
• custom property values should no longer be cached (in the constructor, validate etc.). Read them in-place when you need them!
• Custom skills are now identified as such and copied to /custom_skills in your %APPDATA% directory, outside of the versioned structure.

Skill format changes
If you have custom Skills, you need to update default_config.yaml:

• category (string) is now tags (list of strings)
• Simplified examples to only contain question strings (no more answers)
• there are a couple of new properties for your default_config: platforms, discovery_keywords, discoverable_by_default, display_name etc.
• Removed old migrations from previous versions. If you're upgrading from 1.7.0 or earlier, start fresh and manually migrate some settings!

Known Issues

• Elevenlabs continues to have strict rate limits. If fetching voices and models too quickly, requests may fail. Wait a few seconds between requests and use the wrench icon to retry if needed.
• You still have to login every 24hrs. Sorry, we really tried. Azure things…

Features

• Remote clients (experimental): You can finally connect to your Wingman AI host from other devices on your network. Use your iPad or Laptop as a second screen! While you cannot record your voice using external clients, the TTS output will be played on both your host and external clients. You can also send text from external clients to the host and most user interface interactions will work on external clients, too. Remember that Wingman AI is still running on your PC, so changes are applied there. To connect, simply start Wingman AI on your main PC and then navigate to http://<host-ip>:5173 in the browser on your external device(s). You might have to allow your firewall on the host to listen to this port. You can find out the local host IP by typing ipconfig in a shell or by inspecting your network adapter in Windows.
• Wingman Pro: We improved the performance of our existing regions and added a new Asia region. You have to update to 1.8.0 to get the full benefits, especially if you're in the US or in Asia.
• We added a new Shut up key which defaults to Shift+y (can be rebound in Settings) and cancels any current TTS output via shortcut
• Improved Skills UI that tells you better what Skills actually do. Our in-client tutorial videos are now played in a modal dialog to de-clutter the interface
• We now support OpenAI's new TTS API with our new “OpenAI-compatible TTS" provider and you can use it to connect to local TTS systems like CoquiTTS (xtts2) or others. This is huge because now you can go "full local" for the first time ever and even have voice cloning. If you want to dive into local TTS, we highly recommend @teddybear082's Wingman AI version of CoquiTTS. He has Wingman-specific instructions in this repository and can help you to set it up.
• Our Google provider now supports function calling, so that you can use the gemini-models to their full capabilities. They’re really fast and have quite the “personality”, so try them out! You need an API key but the free tiers are very generous.
• Added Hume AI as new TTS provider. It’s a service similar to Elevenlabs but more affordable.
• Added speed and model parameters and ballad voice for OpenAI TTS
• updated Perplexity models
• we added a {SKIP-TTS} phrase. If you prompt the LLM to add this string to its reply (or write a Skill that does it), the reply will appear in the Wingman UI but not be sent to the TTS provider. This is useful for “silent replies”.
• Skills now have a new on_play_to_user hook that is called just before we send the text to the TTS provider. You can modify the text here to remove smileys from the response, add the {SKIP-TTS} phrase etc.
• You can set default hotwords for FasterWhisper using Settings > Edit default configuration and these are now separate from the Wingman-specific hotwords. In other words: Wingmen (and Skills like UEXCorp) can now add additional hotwords on top of the default ones properly.
• Added new custom property type audio_device for Skills that will render our audio device picker in the UI
• We are now tracking some anonymous usage data in our client to understand better which features you're (not) using. We do not log your backstories or any user-generated content, it's things like "changed LLM provider to x" or "opened Sound Effects Library". We cannot identify you with this data and we won't sell it. You can opt-out in the Settings view.

New Skills

• Audio Device Changer: Lets you change the desired output device per Wingman. This is useful for VTubers who need each Wingman on a dedicated virtual sound device to display avatars etc.

Skill changes

• Vision AI: improved Skill prompt
• Timer: added silent mode
• American/Euro Truck Simulator: fixed location data
• UEXCorp: minor adjustments and optimizations

Bugfixes

• FasterWhisper no longer crashes when you remove a language set previously
• Fixed Elevenlabs provider not loading models (thanks @lugia19)
• Wingman Pro request errors are now properly sent to the client so that we can display them properly in the UI
• fixed missing translations in French and Spanish
• The initial Setup Wizard now asks for your NVIDIA RTX series and sets FasterWhisper’s computeType to float16 if it’s a 5000 card, preventing a crash
• Clicking on the LLM model dropdown and leaving without selecting a new one no longer crashes Wingman AI Core
• models are not (re)loaded properly after entering a secret
• fixed a bug happening during the migration of Wingmen from 1.7.0 which never modified the FasterWhisper configuration before
• fixed UEXCorp Skill importing false from sympy by mistake
• fixed WebSearch skill still providing category instead the the new tags: list[str] in default_config.yaml
• Tower is no longer reset and reinitialized every 30mins when the Azure login token is refreshed

Breaking changes & migration:

• Wingman AI Core now needs a client to be connected to it. We had issues before where Skills were requesting data using the Wingman Pro API before the user was authenticated, so Skill are now initialized after the user has logged in.
• Unfortunately, we have to reset your FasterWhisper hotwords during migration. So you'll have to enter them again. You can still access your old list of hotwords by looking at your old 1_7_0 configs in your APPDATA directory.
• Custom Skills have to be migrated manually as usual. This time we only changed the structure of the default_config.yaml to display Skills better in the UI: category: str was replaced by tags: list[str]. Added new properties display_name (may contain whitespaces) and author. Simplified examples to only contain questions now (so a list of de/en strings).

Known issues:

• Hume currently doesn't list custom voices which we already reported to them. You can still set one by editing the config manually. Let us know if you need help with that!
• Elevenlabs lowered their request limit significantly. If it fetches your voices and you then open the wrench to fetch the models too quickly, the second request might fail. The workaround is to wait a couple of seconds after the voices have been loaded and then click the wrench. You can toggle the wrench to send the request again. If it fails, the model field will just disappear.

Features

• You can now use joystick and mouse buttons as PTT activation keys. This is one of the most requested features ever and we were finally able to pull it off. While we can detect that you pressed a joystick button, we still can’t simulate pressing it.
• We added a new Initial Setup Wizard that will walk you through the most important settings and configuration parameters we always had to explain to everyone: Speaking to your Wingman in another language, running STT on your GPU with CUDA and selecting the nearest Wingman Pro region.
• We built an interactive help aka Tour Guide that you can start from the Terminal view or in a Wingman configuration. It will detect what’s on your screen and show important explanations and hints for several options in the UI. This replaces many of the text blocks we had in the “help “column” before, making the UI much cleaner. The help column will now mostly be used to display short tutorial videos for the appropriate sections. We haven’t produced all of them yet but we can patch them in later without updating the Wingman AI client, so they’ll appear when they’re ready.
• Faster Whisper replaces whispercpp as our default STT provider. It’s significantly faster running on CPUs (the default) and equally fast on NVIDIA GPUs. FasterWhisper supports “hotwords” which you can use to force transcribing complicated names etc. correctly. If you define “ArcCorp” as hotword, it will transcribe everything you say that is similar enough exactly like you typed it out. The names of your Wingmen are always added as hotwords which makes Voice Activation much more reliable now. FasterWhisper allows you to download official models from HuggingFace automatically and this is completely integrated into Wingman AI. If you select a new model from the list, it will be downloaded and can be used directly.You can still use whispercpp if you want to but we removed the code to auto-start/close it, so you’ll have to run it on your own now. You can connect it in the Settings view. It is disabled by default now and will not show up in the list of available STT providers unless you enable it in the Settings (similar to XVASynth).
• The Terminal now shows better benchmark results per message. You can hover the processing time to learn how long each step of the processing took. This will make it much easier to find out where potential bottlenecks in your configuration are.
• Spanish and French were added as client languages. We obviously used Wingman AI to translate itself, so please tell us if the translations suck and we will change them.
• We removed the ‘Advanced’ tab in the Wingman configuration and now allow all users (including those in the free plan) to change all settings that were under Advanced before. This means you can now change STT, TTS and AI providers as you wish and use your own API keys without having to touch any configs manually. Editing Commands and Skills in the UI will remain a Pro-only feature.
• The system prompt where your backstory is inserted as variable can now be overwritten per Wingman, similar to most other configuration parameters. It will appear next to the backstory in the UI.
• You can now define stop and resume actions for audio playback Commands, meaning you can start playing a sound file in one Command and then stop or resume the playback in another one.
• added a new option to Settings that displays incoming Wingman messages larger and fades out the background - we call it Zoom View. Some love it, others don't, so it's disabled by default.
• Wingman Pro models are now loaded via an API in our backend so that we can update them without having to update the Wingman AI client.

Skill changes

• StarHead has new features: Get trading information of specific shop - you can now ask if a commodity can be traded at a specific shop and what the price is. Get trading shop information of celestial object - you can now ask for shops trading a specific commodity and what the price is.
• UEXCorp was completely reworked - big thanks to @Lazarr for testing! Receiving item data or trade routes can now be toggled on/off individually. Better names/terms matching with dynamic AI suggestions. The AI now understands relations between different objects better. Automatic FasterWhisper hotword list export to your Wingman. This is huge and means your Wingman will know all the correct spellings for SC terms and will transcribe them much more reliably. Data is now kept up-to-date during a session automatically. Getter profit calculation. Removed option to get additional lore information for ships and locations (for now).
• FileManager can now handle .zip files (extract & compress)
• Spotify is less likely to crash Wingman AI if it's not setup correctly. We'll also make a video explaining how to set it up soon.
• Most Skills got minor updates & bugfixes

New Skills

• Time and Date lets your Wingman know… just that. This is apparently useful in certain scenarios and was a requested feature by the community.

Bugfixes

• fixed a small bug with great effect that would let your voice input appear in the UI with a delay - only after the AI response was generated. This only happened in Focus mode (if a specific Wingman was selected in the Terminal view) and let people think that their STT provider was working too slowly.
• Wingman AI Core now logs more reliably, even when it crashes early. This means you have to start it manually from the command prompt less often to see what’s (not) happening.
• fixed OpenRouter suddenly throwing errors when you send tools to a model that doesn’t support them. We also display if the selected model can call tools or not in the UI now.
• fixed an issue when trying to stop audio playback
• fixed ‘Right Alt’ modifier key and some other extended Windows key codes not being recorded correctly
• Wingman AI Core is now running on port 49111 by default (which is widely unused) and the port can be changed using a command line parameter or by the Tauri client
• Wingman AI Core now also detects Skills that are only present in the AppData directory but not in the source directory
• resolved Wingman AI Core model naming convention warnings on start (coming from Pydantic)
• migrated to Tauri v2 which was a pain to do but is important for future updates and maintenance.
• Deleting a Wingman is now guarded by a text input prompt so that you can’t delete them by mistake, e.g. when editing Commands
• fixed special characters used in Commands crashing Wingman AI. They are now sluggified properly. We also fixed the regex validating Wingman names leading to the same error.
• We tried everything to require you to login less often. The current state is that you will always stay logged in as long as Wingman AI is running. If you exit Wingman AI, you are required to login again after 24 hours of inactivity. This limitation is set by Microsoft/Azure and we cannot change it.
• We made “AI filler responses” a Pro feature because it will throw “You need a subscription” errors if Wingman Pro is selected as provider but you don’t have a Pro subscription.
• The text input now regains focus after it was disabled/enabled during TTS output. This means you can keep typing without having to click on the input again after each message.
• Various small UI fixes and improvements

Skill Changes

• ImageGeneration can now save to file and passes the file path to chained function calls

Bugfixes

• Closing Wingman AI client now properly terminates whispercpp and xvysynth processes (if it launched them)
• fixed config directories being recreated when the default config dir was changed (on Windows)
• fixed "white screen error" when changing the default Wingman in the active config using the client
• Wingman Pro: hide "Subscribe" buttons if you already have a subscription
• fixed some minor client issues

Features

• We were forced to implement a **new payment provider **for Wingman Pro. See our announcements on various channels for more information.
• Audio Library: There is now a new audio_library directory in your config directory and you can put any number of .mp3 and/or .wav files in there. You can also structure them im subdirectories. These files are imported into Wingman AI and you can use them as new action type in Commands or in Skills. You can now create commands that press hotkeys and then play a sound file - or the other way around. If you select multiple files, a random one will be chosen to play. Sounds familiar, HCS Voice Packs/VA?
• AI Sound effect generation: You don’t have an audio library yet or need a new sound for a very specific use case? Use Elevenlabs to generate mp3 files on the fly using a simple prompt!
• We redesigned Command Configuration. It’s no longer modal but a dedicated view giving it more screen space. We also optimised the UX and made all the options for instant activation and responses much easier to understand. You can now also tell the AI not to say anything after executing command which is useful if you just want to play a sound file, for example.
• Wingman Client now has a new Audio Library drawer that lets you see your current sound effects and Elevenlabs subscription details (including characters left)
• Added new LLM providers & models:
• Cerebras: A very(!) fast provider similar to Groq. They offer access to llama-3.1 models and are currently free-to-use. New models will be added automatically to Wingman AI.
• Perplexity: Trusted real-time data powered by llama-3.1. Does not support function calling in Wingman AI but there is a Skill to solve that (see below). It’s a paid service, you will need a funded account with an active API key
• OpenAI models are now fetched dynamically so that you can use all available models with your own API key
• Config migrations are now sequential, meaning that you will always be able to migrate from any(!) version since 1.4.0 to the new version, even if you skipped some in between. Note that custom (non-official) Skills currently can’t be migrated and Wingmen that are using these, will fail to migrate. Just delete the Skills from your Wingman configs manually, then copy over your custom Skill code to 1.6.0 and run the migration again. Finally re-attach your custom Skills to your now migrated Wingmen in the client.

Skill changes

• FileManager can now read PDF files

New Skills

• Image Generation: Let your Wingman create images on the fly using Dall-E 3 (free & unlimited with Wingman Pro)
• MSFS 2020 Control: Control and retrieve data from Microsoft Flight Simulator 2020 using SimConnect
• ThinkingSound: Use files from the new AudioLibrary to play sounds while the AI is generating responses
• AskPerplexity: Use the Perplexity API to get up-to-date information on a wide range of topics. The Skill allows you to run a model with function calling capabilities (like Wingman Pro/gpt-4o) as main driver for your Wingman while still providing access to recent data from Perplexity.

Bugfixes

• Fixed secret prompts immediately disappearing before you could enter anything. You also no longer have to restart Wingman AI after entering certain secrets for providers or Skills
• Fixed a bug where trying to remove a voice from the VoiceSelection list (used in RadioChatter and VoiceChanger) removed the voice before that
• Better error handling and performance improvements for Elevenlabs
• The “ToC” modal no longer auto-closes before you can click it
• Closing Wingman Client now properly unloads Skills and ends the XVASynth task (instead of accidentally launching it)
• XVASynth: added missing “disabled” label
• Fixed Discord invite links (because our old one expired) & updated some docs

Features

• Autosave: You complained in our survey that the Wingman AI client was unresponsive and sometimes felt sluggish and as your favourite UX-driven development company, we listened. From now on, almost any action and config change you can make happens ad-hoc and no longer requires reloading the config you’re in. This brings a couple of benefits: increased performance as loading a config is expensive, way less conversation history resets, no more “Save” buttons, more real estate in the UI as we don't need the ActionBar anymore and you can't forget to save.
• Config Migrations. Finally. Wingman AI 1.5.0 (and future versions) will now take your configs from previous versions and migrate them to the current format, meaning that you’ll keep almost all of your old secrets, settings, configs and Wingmen. You’ll find a new .migration file in your 1.5.0 config directory after the first start. This file is a log and as long it’s there, Wingman AI will not attempt to migrate again on start. If you want to re-migrate your 1.4.0 configs for some reason, delete this .migration file (and the 1.5.0 configs), then restart Wingman AI. We only support migrating from 1.4.0 to 1.5.0 but future versions might also be able to migrate step-by-step starting from version 1.4.0. If you are a developer and have built your own skills in older versions, you have to copy the files from /[old_version]/skills/[your-skill] to /[new_version]/skills/[your-skill] manually because we cannot migrate unknown skills reliably. Restart Wingman AI afterwards.
• Our beloved TTS provider whispercpp is now bundled (and auto-started) with Wingman AI and is the new default for everything so that everyone can use it. Check out the new section the Settings view to set it up. If you have a NVIDIA RTX GPU, make sure to check the “use CUDA” option. If you’re using custom models (other than ggml-base.bin), just copy them into the new whispercpp-models directory in your Wingman AI installation directory and the UI will find them.
• Edit default config: In the Settings, you’ll find a new button and view that lets you change the defaults config (defaults.yaml). This includes the highest level system_prompt that Wingman AI uses. Be careful if you modify this and please only do so if you know what you’re doing. Yes, editing the default config also has autosave.
• Added volume slider to Wingman configuration. If it’s set to zero, TTS processing will be skipped entirely.
• added OpenAI’s much cheaper and faster gpt-4o-mini model and made it the new default. The larger gpt-4o model is still available in case you need it.
• added Google Gemini as LLM provider. Function calling doesn’t work (yet) but you can use it for chit-chat and roleplaying. It’s very fast and pretty good - try it! Unfortunately, we can’t provide it with Wingman Pro, so you’ll need your own API key. The smaller model is currently free.
• Groq and Elevenlabs models are now fetched using their API so that new models will be available in Wingman AI immediately and without an update. Check out Groq’s new llama3.1 models with function calls and Elevenlabs’ brand-new eleven-turbo-2.5 model finally supporting multiple languages.
• We removed the summarize_provider so that tool calls can now be chained, meaning that the response to a function call can call another function and so on. This doesn’t sound like much but is actually huge. You (and skills) can now basically give commands like “Use the XYZ skill and then do ABC with the result”.
• We split settings and configuration for XVASynth and improved its UI support meaning that it has nice dropdown for all your downloaded models and voices now. If you want to use XVASynth, install it via Steam, then enable and configure it in the Wingman AI Settings view. After that, it will appear in the list of TTS providers. We cannot migrate your old XVA settings, so please do that once, even if you were already using XVASynth.
• Wingman AI Core can now open directories and files using Windows Explorer or OSX Finder. We added a new useful buttons like “open logs” or “open config directory” to the client.
• We added new custom property types "VoiceSelection" and "Slider" so that skills can display our fancy UI components for them. No more writing JSON into text boxes to select voices...
• We removed DirectSound drivers for audio devices because it caused more problems than it solved. The only available driver is MME now. Therefore, we have to reset your audio device settings during the migration, sorry. Make sure you check the “Settings” view after upgrading.

Bugfixes

• Links in LLM responses now have target="_blank" set so that they’ll open in a new browser Window
• Skills are now unloaded correctly when removed. No more infinite zombie timers. We also added some new hooks for developers.
• fixed an issue with our AudioPlayer that sometimes resulted in a crash when using the RADIO_LOW or RADIO_MEDIUM sound effects
• fixed an issue with mic selection while using VA
• fixed an error occurring when using a default audio device
• fixed renamed configs being recreated on startup (again)
• fixed not being able to un-default renamed Wingmen
• fixed Wingmen being disabled when a secret from their configuration was missing on startup
• fixed Elevenlabs get_available_voices API endpoint changing for no apparent reason and without warning us ;) Thanks again to @lugia19 who fixed this very quickly in his great elevenlabslib
• Request a feature is now opened in a new browser window because integrating it could cause a “blank screen” error.
• If Wingman AI Core fails to start, we now show better trouble-shooting infos pointing to our Discord #support channel in the client.

New Skills

• APIRequest: This one is a powerhouse. Point it to a docs page or (OpenAPI) spec for an API of your choice and it will understand the endpoints and then be able to call them - all on-thy-fly! We already gave it the Wingman AI Core API spec and well.. it was able to “send text” to another Wingman immediately without us ever implementing that. Wingmanception! You can also feed it any public online API like Pokedex or whatever. Try it out!
• RadioChatter plays customizable random AI-generated chatter over time. You can even answer to these messages if you want to. It also uses our new custom property types now (in case you were already using the Discord versions).
• Auto-Screenshot: Takes a screenshot if you request it or in moments where your voice input suggests interesting, scary or funny moments. Example: “Aww look, a spaceship!” => takes a screenshot
• ATSTelemetry: Retrieve game state information from American Truck Simulator or Euro Truck Simulator 2
• NMSAssistant: fetch information about No Man's Sky items, elements, crafting, cooking, expeditions, community missions, game news, and patch notes. Powered by NMS Assistant API.

Skill changes

• ControlWindows and FileManager got some new capabilities like new supported file formats, text-to-clipboard and appending text to existing files
• Timer skill can now loop and works more efficiently
• VoiceChanger now makes use of the new custom property types and offers a nice voice selection in the UI
• Vision now outputs the direct LLM responses in debug_mode

Skills:

• added Timer: Let your Wingman delay and time command executions or responses
• added VoiceChanger: Let your Wingman change its voice automatically and customisable
• added WebSearch: Searches the web using the DuckDuckGo search engine. You can also specify a specific page/URL that it will crawl.
• added TypingAssistant: Let your Wingman type text for you (in any app etc.)
• added QuickCommands: Learn command phrases you use regularly and auto-convert them to instant activation commands to execute them faster
• added Vision: Use AI to take and analyse a live screenshot, then ask questions about it
• added FileHandler: Read and create files and directories
• UEXCorp now supports even more UEX API v2 features and no longer requires an API key
• ControlWindows can now use your clipboard and supports window snapping

Features & Changes

• added Clippy Wingman - the famous assistant we all know and hate, now resurrected with AI powers and using a lot of our new skills. He resides in a new config name “General”.
• added AI avatar generation for your Wingman using DALL-E
• improved the quality of all of our audio effects a lot
• “ROBOT” sound effect is now called “AI”
• “RADIO” sound effect now has 3 different variants to simulate a low/medium/high quality radio device
• new Apollo Beep sound
• Your Wingman can now say instant filler responses (like “Give me a second to…”) for long-running tasks
• The Wingman AI Client can now render Markdown responses with links, images, different text formats etc. (like SlickGPT)
• you can now define multiple commands with the same instant activation phrase and they’ll all be executed (with a better response) if you say the phrase
• Wingman AI client no longer saves your entire Wingman configuration but only the differences to the defaults.yaml config. This is a huge change and will help us with future config migrations and allow us to finally make the defaults editable in the UI in a later version.

Bugfixes

• fixed renamed or deleted default Configs/Wingmen being recreated on restart
• fixed a bug enabling Voice Activation even if it was disabled in the Settings if you (accidentally) pressed the un/mute key (default: Shift+x)
• fixed a bug causing dependencies defined by Skills not being loaded properly
• fixed drag&drop to reorder action “rows” in the Commands tab
• fixed silent error/crash on start when no mic is connected
• added missing Azure STT language selection for Voice Activation (or PTT with external clients) in Settings
• fixed creating new configs from templates ("Empty" vs. "Star Citizen")
• fixed conversation not being cleared if the "clear button" was clicked in "View all" terminal mode
• fixed mouse button actions not being displayed in the Command list

• added the updated UEXCorp skill that now works with the new UEX API v2.0
• Free users can now fully explore (but not save) all skills in the client
• StarHead skill is now attached to the Computer Wingman by default so that Free users can use it
• improved default contexts of our demo Wingmen
• fixed a bug with un/muting that lead to Voice Activation being enabled when it shouldn't
• fixed an issue when handling Alt Gr in command bindings
• fixed renamed default Wingmen being recreated on restart
• fixed a bug caused by Wingmen without configured activation keys
• fixed a bug with saving the Basic configuration
• gpt-4o for Wingman Pro/Ultra is now served via Azure (instead of OpenAI)

• Added gpt-4o model. The latest & greatest from OpenAI ticks all the boxes for Wingman AI, so we are making it the new default model for... everything. We've even removed the gpt-3.5-turbo and gpt-4-turbo options because we're sure we won't need them anymore. General responses should already be significantly faster, and it will get much faster once OpenAI unlocks audio input/output to the API. The model also seems to work well with our new contexts and capabilities. Please let us know (with examples) if we are wrong!
• StarHead is now a Skill, meaning you can attach it to any Wingman.
• Added new Control Windows Skill to launch applications, minimise/maximise Windows etc. For example, tell it to “Open Spotify”. Note that the skill currently only looks for applications in your Start Menu directory (%APPDATA%/Roaming/Microsoft/Windows/Start Menu/Programs), so if it tells you that it cannot find an application, create a shortcut in that directory.
• Added new Spotify Skill to play songs or playlists, "like” songs, control the playback device and more. The skill is very intelligent and honestly takes Spotify to another level because you can ask it things like: “Play the theme song from Bad Boys”. Try it out, it’s really good! You will need to create a developer account for your Spotify Premium account and then connect Wingman to it. The Spotify client has to be running on any one of your clients and must be playing a song. Otherwise, the Spotify Web API won't find any active device for your account. You can find more information and instructions on our Discord server.
• Added Groq and OpenRouter (in favour of the old "Llama") as AI providers. You can select any of their models from a list in the UI, but only Mistral and Llama3 have been tested with AI function calls. However, you can use other models for roleplaying etc. and experiment with them. Our demo contexts may need to be tweaked to work properly with other models.
• Added "Local LLM" as an AI provider. This makes it easier for you to connect to different models from different providers. Use Local LLM instead of the old OpenAI > Advanced > base_path setting to connect to Ollama, LMStudio etc. You can now enter your own custom model name and are no longer limited to the drop down choices in the client.
• Updated default commands for Star Citizen 3.23. There have been a lot of changes in the last SC patch and our commands are by no means a complete mapping for everything. Feel free to improve these bindings and post your updates in Discord. We can collect your changes and then release them for everyone to use.
• Added a handler and websocket allowing you to connect esp32 devices to Wingman AI. This is what we used in our Walkie Talkie for Wingman AI short video. We’ll publish a blog post about this but until then, contact @Timo if you’re interested in connecting external devices to Wingman AI using OpenInterpreter etc.
• Added Debug Mode toggle to global settings. Enable this for more verbose system messages and additional benchmarks-
• @JayMatthew’s beloved UEXCorp Wingman has already been converted to be a skill, too. Unfortunately, the UEXCorp API was changed completely yesterday, so he will fix the skill as soon as possible and you can then install it (without another Wingman AI update) when it’s done. You can even use it together with StarHead now! You may need to give special commands to determine which skill you want to prefer when asking for a good trading route. We are experimenting with this, but we need more data and experience from you, so please let us know how it goes.
• fixed crashes when AI function calling with llama3 models

SlickGPT Pro (For Wingman Ultra users):

• added gpt-4o model for both SlickGPT Free and Pro users
• Free users can now use llama3 (for free via Groq) and mistral-large with their own API keys. We are waiting for a fix in Azure to serve these two options in our SlickGPT Pro backend.
• various UI improvements on desktop & mobile viewports

• We updated the StarHead Wingman to support better trading data and added a new function to get detailed ship data. You can ask for ship details or even compare different ships or ship systems. Try something like "What is the difference between a Constellation Taurus and a Constellation Andromeda", for example.
• We worked around the super slow performance of GPT-4-Turbo, so you can now use it again. It's still slower than GPT-3.5-Tubo.
• Remove emotes from Audio Playback, if they are in the text result. Sometimes the AI want's to express emotions like this, but it was very annoying in the audio output.

• Mistral AI has been added as a top class provider. You can use your own API key to access all their cloud models or use the most powerful one (mistral-large) with Wingman Pro. Please note that if you use Mistral, you will need to adapt your contexts, as it does not seem to handle our examples written for GPT very well. Mistral supports function calls, so start with "You are a helpful assistant", which should work with AI commands, and iterate from there.
• We have also added experimental support for Llama3. It's all ready in Wingman AI, but not yet in Azure, so if you try to use it with Wingman Pro, your requests will time out. We are waiting for Microsoft to fix this and will let you know when it works (hopefully with no additional client update). Until then, you can use 3rd party providers like Groq (super fast!) or host it locally and play around with it. Function calls won't work yet because we need to write some more code to parse them, but you can already chat with it and run instant commands as usual.
• Reworked command UI. You can now add mouse and write (text) actions. You can also delete individual actions from a command's action list and reorder them using drag and drop.
• Improved Windows audio device settings. Wingman AI should now handle disconnected devices or changed hardware configurations much better.
• Improved function call parsing to be more robust when using local LLMs.
• Added a UI option to make OpenAI the default provider for all Wingmen settings. This makes it easier to use Wingman AI without a subscription. Note that you can still manually change your default provider in the config defaults.yaml in your config directory.

Unfortunately we have to change our payment provider for tax reasons. We are switching from Stripe to Paddle. We have already changed the whole process for new users/subscribers. If you have a subscription with Stripe, you can still manage it using the client and our website. We are currently trying to determine if we need to cancel and migrate current Stripe subscriptions.
If you would like to help with the transition, please cancel in Stripe and then re-subscribe with Paddle if your subscription has expired. You don't need to do anything at the moment, but if this changes we'll let you know!

• fixed error when switching config
• text input is now disabled if the focused Wingman is disabled

• You can now send text messages to your Wingmen! To do so:
• Select a Wingman in the bar on the left to open a filtered "Focus view" which only shows the conversation history with this particular Wingman. If you talk to another Wingman while in this mode, the view will automatically switch to the activated Wingman. Select "View all" to see the shared "Terminal" with messages of all your Wingmen (as usual).
• fixed handling of "extended" keys and added a default "hold time" of 0.1s which should greatly improve keypress command executions. If you had "Wingman doesn't press the key" issues before, please try again with this version!
• added "extended macro" recordings, meaning that key down/up events will be recorded separately. This gives you more flexibility to record complex and staggered macros
• The client now has a button to reset the conversation history (in both Focus and Terminal view)
• reworked "energy threshold" controls for Voice Activation in the client. It's no longer a slider but a numeric input with presets (low/standard/custom) and better hint texts to explain what meaningful values are.
• fixed Microsoft password policy to be less restrictive
• fixed Conversation/Summarize model not being saved properly when using the client
• changed/moved various API endpoint URLs

• added user authentication (to the client)
• added Wingman Pro as new default provider with several choices for you: STT: Azure Speech or OpenAI Whisper (via Azure); TTS: Azure TTS or OpenAI TTS (via Azure); AI Conversation & Summarization: gpt-3.5-turbo or gpt-4-turbo (via Azure); Voice Activation: A special implementation of Azure Speech STT; Regions: Europe or USA - pick the nearest for best latency!
• ElevenLabs: output streaming can now be disabled; lowered default latencyOptimization (from 3 to 2) to prevent choppy audio
• added OpenAI base_url and organization params to the UI, allowing you to plug in a local LLM as stand-in replacement (or a proxy) easily
• added a lot of hint texts to the UI & updated docs
• migrated to new elevenlabslib version and updated all important packages
• stop current playback if a new one starts, preventing "overlapping" audio
• Voice Activation now prefilters all recordings and checks if they contain frequency bands resembling human voice
• print version number and update notification on start
• fixed some keybinding issues with "special” keys like Alt and Alt Gr
• fixed Star Citizen "SelfDestruct" command not being executed due to moderation policies
• fixed a crash occurring on command execution
• fixed Azure speech languages not being passed correctly
• fixed a performance problem with Voice Activation
• fixed Wingman initialization errors not showing up in the UI
• fixed a problem causing slowdowns if whispercpp was configured but not running
• fixed a bug preventing to change the AI model to gpt-4-turbo when using Azure
• fixed StarHead not setting gpt-4-turbo as default model correctly

Known issues:

• gpt-4-turbo via Azure/Wingman Pro is VERY slow at the moment. Microsoft is informed and (hopefully) working on it.
• Voice Activation sends way too many blank messages to the STT provider. We found a cool solution that we will test until the final release but it’s not in the RC yet.
• If you only see a blank white screen after starting the client: Close Wingman AI. Delete this directory: %APPDATA%/Local/com.wingman.dev. Restart Wingman AI.

Happy Easter holidays and have fun!