Release notes
Unclaw updates itself. When a new version is out, the app fetches it in the background and applies it on the next launch, so there is nothing to download or reinstall.
Newest first. Version numbers are shared across platforms, but a release sometimes reaches one platform a few days before another.
August 27, 2026
Mac first, with Windows and Linux following. This one includes an updated character engine (about a 3.4 GB background download). Models and voices are unchanged, so nothing else big to fetch.
The efficiency numbers below compare against 1.1.3, the last release we announced, so they cover everything since.
- The stutter is gone. Since the beginning, the stream had a small hitch at a steady rhythm, about once a second, subtle enough that every frame-rate counter swore things were fine. The cause turned out to be the video encoder quietly inserting a full keyframe every 1.2 seconds, and each one arrived late. Measured before and after with presentation-level timing: 42 visible hitches per two-and-a-half minutes before, zero after. This fix needs the updated engine, which is why this release carries one.
- Lighter on your machine. Around 30% less CPU and about 5% less GPU than 1.1.3, with roughly 400 MB less video memory. Most of that comes from the display path: the character’s frames now go straight to the screen instead of being encoded and decoded like a video stream, which also cuts latency and sharpens the picture.
- A Chrome extension. Your companion in a browser side panel, where she can see the page you are reading, answer out loud, and act in your browser when you ask. Approved and available on the Chrome Web Store. See Chrome extension.
- Pocket is the default voice now. The audio problems that got it pulled in 1.1.9 were a state-handling bug in how sentences were generated back to back, not the voice model itself: the second sentence onward was being conditioned on leftovers of the first. Fixed and verified word for word. It is also louder now (it used to sit noticeably below the other engines), and it is the fastest voice we ship: speech starts about a hundredth of a second after a sentence is ready, entirely on your machine, no key needed. Your saved choice is respected; only fresh installs default to Pocket. Kokoro and Qwen3 have been retired, and anyone using them moves to Pocket.
- A new onboarding experience. Setup now asks who you are before it asks for an account, sign-in moved to the end, and your companion greets you live rather than from a recording: she uses your name, the actual day and time, and varies what she says.
- More expressive while she talks. Expressions carry through a reply instead of being flattened by the mouth movement.
- Phone calls connect more reliably. A stale session on the Mac could make your phone show the Mac as available while calls silently went nowhere. Several causes fixed on the Mac side, including leftover relay processes from previous runs interfering with new calls.
- Characters keep their own identity everywhere. Chats from the browser extension and the phone used to answer with the default persona and voice regardless of who was on stage. The character you see is now the character who answers, everywhere.
- Signing out resets properly. The app used to keep the previous account’s character setup on screen until you refreshed, and signing back in did not re-apply your saved look without a restart. Both fixed: signed out you get a stock character, and signing in restores your own.
- Memory use is down when Pocket is active: about 800 MB returned by tightening the voice engine’s working cache, with no audible cost.
There is now a documentation site at docs.unclaw.io, including a guide for the Chrome extension.
When the Windows build lands it adds DLSS 4.5 support.
August 20, 2026
Mac only, and a small one: same character engine and models as 1.1.8, so there’s nothing large to download this time.
- Fixed the character looking zoomed in on first launch. On a cold start the engine takes a moment to come up, and the request telling it what resolution to render could be sent before it was listening. Nothing retried, so it stayed at a small render size that got stretched to fit your window until you refreshed. It now checks what’s actually being rendered against what it asked for, and re-sends if they disagree.
- Fixed washed-out colour on first launch, which had the same root cause in a different place: the colour-matching filter could be applied before the page had finished setting it up, and silently did nothing. It now retries until it takes.
- Pocket has been temporarily removed from the voice engines. It was fast, but the audio quality wasn’t good enough to keep in front of you. If you had it selected, you’ll be moved to Supertonic-3, which is also local, needs no key, and has real cloned voices for every character. Pocket will come back when it sounds right.
August 20, 2026
Mac only. This one includes a new character engine, so expect a larger download than usual: about 3.5 GB for the engine and 1 GB of models and voices. It happens once, in the background, and the app tells you when it’s ready.
- A new local voice: Pocket. Cloned per character, runs entirely on your machine, and it’s the fastest engine we ship, generating speech around 17 times faster than real time. Every character has its own cloned Pocket voice, including the ones you’ve unlocked.
- A sharper picture. The character now renders at your window’s true resolution instead of being drawn small and stretched to fit. Most noticeable on larger windows.
- A new display path. On Apple Silicon the character’s frames now go straight to the screen without being encoded and decoded first. Lower CPU use, lower latency, and the blur and glass effects behave properly again.
- Steadier motion. The engine now runs at a fixed frame rate, which removes the uneven pacing some conversations had.
- Reverted an animation optimization that made body movement look laggy.
Known issue: photo-to-character is still turned off while the last of its infrastructure lands.
August 19, 2026
Mac only. Windows and Linux stay on 1.1.6 for now.
- Expressions work while your companion is speaking. Surprise, delight and concern used to be flattened the moment speech started, so replies came out with a still face. They now play through speech at full strength.
- Fixed the grin flash. A large, wrong smile would appear for a fraction of a second before certain expressions. Gone.
- More human idle behaviour: blinks land on sentence boundaries, the gaze wanders and returns the way a person’s does in conversation, and head motion follows the melody of the voice rather than the volume.
- Closing the window actually closes everything. The red button used to leave Unclaw sitting in the Dock with the character engine still running in the background, burning battery until you force-quit it.
- Fixed the error dialog when quitting, and a hang that could leave the app unable to finish shutting down.
- Custom characters now keep the hair and eye colour you picked. Every custom character was silently falling back to default grooming.
- The window no longer floats above everything by default. Use the pin in the title bar when you want it on top.
Known issues: photo-to-character is turned off in this release while the last of its infrastructure lands. Pocket, the new local voice engine, is not yet selectable, for the same reason.
August 5, 2026
- Character downloads are now verified byte for byte before they are used, and a half-written or stale character file is quarantined instead of handed to the engine. Fixes characters that failed to appear after a purchase or an interrupted download.
August 5, 2026
- New character engine build. Video capture is now zero-copy, which takes a meaningful bite out of CPU usage during a conversation.
- The backdrop is no longer lit, cutting GPU load by around 22% with no visible difference.
- Store characters were re-cut against the same engine build, so a purchased character always matches the engine that loads it.
August 5, 2026
- Smaller install. The old neural expression model was retired in favour of the captured expression engine, removing about 405 MB from the download.
- The mood and expression library now ships as data, so expressions keep improving without a new engine build.
July 21, 2026
Shown at SIGGRAPH, and shaped by the feedback there.
- Unclaw CLI.
npx -y unclaw-cli@latestconnects your existing agentic coding tools (Claude Code, Codex, OpenClaw) to Unclaw, so your companion talks to you while you work. - Body proportions. Characters support height and weight changes. Available on Grace and Kevin to start.
- Much lighter. The character engine dropped from 10 to 11 GB of unified memory down to 5.5 GB on Mac, and the AI runtime from over 3 GB to about 700 MB, with no loss of visual quality.
- Custom backgrounds, including two animated ones (ASCII and Cubes).
- More customization: eyebrows and eyelashes, 34 hairstyles, 20+ eyebrow styles.
- Camera presets to flip instantly between close, medium and far.
- New lip-sync model with combined expression handling. Stronger, more accurate mouth shapes.
- New gestures the character chooses for itself: talking, explaining, arguing, and more.
- New speech recognition model (Moonshine), faster and lighter.
- Kevin, a new free character.
July 20, 2026
- Customization polish and voice pipeline fixes.
1.0.12
Section titled “1.0.12”June 8, 2026
- Multiple characters. Add as many companions as you want and customize each one independently. Only one is loaded at a time, so memory use stays flat no matter how large your roster gets.
- Claws, an in-app currency you earn by talking with your companions and spend on new characters and items. It cannot be bought.
- Supertonic, a new local voice engine. Small, fast, and good, made to feel like it streams through chunking.
- Custom colours on clothing items.
- New Unclaw logo.
- About 25% less video memory used on Mac.
- Updater issues resolved.
1.0.10
Section titled “1.0.10”May 27, 2026
- Automatic updates. From this version on, Unclaw updates itself. No more manual reinstalls.
- Many more AI providers, including using your existing Claude Code, Codex or Gemini subscription through their CLI. If a tool is not installed, Unclaw walks you through it. It is now possible to use Unclaw without ever touching an API key.
- Resolution follows your window, so a small sidebar window renders a small area. Quality presets (low, medium, full) added.
- Dynamic ports. Unclaw picks free ports at startup, so it can no longer fail to launch because something else holds a port it wanted.
- Celebrations. Your companion celebrates your wins, or use
/celebrate. - Settings moved under the profile icon and given a cleanup.
May 25, 2026
- Fixed the Mac black screen. The cause was Unreal Engine’s own pixel streaming, which carries years-old bugs on Mac. Unclaw now uses a Mac-native streaming backend written for this purpose, and open-sourced.
Unclaw for Mac
Section titled “Unclaw for Mac”May 23, 2026
The first Mac release, after a significant refactor. Everything from the lip-sync model to the end-user experience differs from the Windows build.
- One app. Unclaw and its AI runtime install together as a single application rather than two separate installs.
- Customization widget for clothes, hair, footwear, lighting angle and lighting colour.
- Significantly improved and optimized graphics, with much nicer lighting.
Known issues at the time: lip-sync quality behind the Windows build, occasional character engine crashes with immediate respawn, and the expression model largely non-functional on Mac. All since addressed.
May 13, 2026
- Fixed the double audio bug.
- Fixed the close button not actually closing Unclaw.
- Fixed the Qwen3 local voice install.
- Fixed voice mode.
- Agentic mode for local models, best on Gemma 4 and Qwen 3.5 / 3.6.
- New abilities: browser navigation, file system access, computer use, and local code execution.
- Weather, stocks and news widgets enabled for anyone with a Gemini key set for search.
- Thinking levels (low, medium, high) for reasoning models.
- Images. Paste a photo into the input bar, or press Control + Shift + G to screenshot, and it goes to your local model.
- Reset button for the sliders.
- Large latency improvements across all models, and better answer quality.
May 2026
The first public test version.