Changelog

What actually shipped.

Latest releases first, fetched from GitHub Releases with a five-minute cache. Open a release for its full notes and downloadable builds.

Cinderpaw v2026.10.28

Skills you install are used without you naming them, and long answers are no longer cut off.

Fixed

Long answers are no longer cut off. Every message carried a hidden limit of about 3,000 words, so a long document written in one go could stop and show "The model returned an empty response". The limit now follows what your AI model can really write.
A skill you installed is used when it fits, without you typing its name. Before, a skill added from a link sat unused unless you named it.
An app or web page Cinderpaw makes runs the moment it is saved. While it is written you see "Building it now" and a line count, instead of raw code that made you wait up to a minute after the app was ready.
Cinderpaw can check what it typed into a post box on X, LinkedIn, Gmail and Slack. Before, it read those boxes back as empty and asked you to check.
A file with spaces in its name is no longer reported missing after Cinderpaw saved it.
The install card says what it installs ("Install hyperframes on this computer?") and shows the full command in its own box.
A long chat stays smooth while a reply is written, up to twice as smooth in chats with hundreds of messages.
Safer zip downloads. The part that unpacks the tools Cinderpaw downloads on Windows moved to a fixed version (RUSTSEC-2026-0367, 0368 and 0369).

New

A skill from a link brings its whole folder of text files, not only SKILL.md, and the card says how many come with it. Programs inside a skill are still left out, and the card names them: Cinderpaw says those steps cannot run here instead of guessing what they would have produced.
The model switcher opens like a Dynamic Island, and cards that wait for you pour in smoothly.
v2026.10.28 on GitHub →

Cinderpaw v2026.10.28-beta.1pre-release

Skills you install are used without you naming them, and long answers are no longer cut off.

Fixed

Long answers are no longer cut off. Every message carried a hidden limit of about 3,000 words, so a long document written in one go could stop and show "The model returned an empty response". The limit now follows what your AI model can really write.
A skill you installed is used when it fits, without you typing its name. Before, a skill added from a link sat unused unless you named it.
An app or web page Cinderpaw makes runs the moment it is saved. While it is written you see "Building it now" and a line count, instead of raw code that made you wait up to a minute after the app was ready.
Cinderpaw can check what it typed into a post box on X, LinkedIn, Gmail and Slack. Before, it read those boxes back as empty and asked you to check.
A file with spaces in its name is no longer reported missing after Cinderpaw saved it.
The install card says what it installs ("Install hyperframes on this computer?") and shows the full command in its own box.
A long chat stays smooth while a reply is written, up to twice as smooth in chats with hundreds of messages.
Safer zip downloads. The part that unpacks the tools Cinderpaw downloads on Windows moved to a fixed version (RUSTSEC-2026-0367, 0368 and 0369).

New

A skill from a link brings its whole folder of text files, not only SKILL.md, and the card says how many come with it. Programs inside a skill are still left out, and the card names them: Cinderpaw says those steps cannot run here instead of guessing what they would have produced.
The model switcher opens like a Dynamic Island, and cards that wait for you pour in smoothly.
v2026.10.28-beta.1 on GitHub →

Cinderpaw v2026.10.26

Documents are written calmly to the end, and Linux updates by itself again.

Fixed

A long document no longer rushes at the end or freezes the panel. Ask for a PDF or any document and you see it written at an even, readable pace, about 110 letters a second, from the first letter to the last, with any AI model: one that writes slowly, one that writes very fast, or one that sends the whole document at once. Before, the end of a long document came in a blink and the panel could stutter.
The writing cursor stays on the last letter you can see, instead of jumping ahead to where the AI had got to.
Linux updates by itself again. If you have 2026.10.25 or older, it updates like on Windows and Mac. The tray icon next to the clock uses libayatana-appindicator3-1: most desktops have it already. Without it, Cinderpaw still works, and closing the window quits it instead of keeping it in the tray.

New

What's new shows this update's biggest changes, once, the first time you open Cinderpaw after updating: tasks for later, page watches, asking before it pays, and bringing your things over from OpenClaw or Hermes.
v2026.10.26 on GitHub →

Cinderpaw v2026.10.26-beta.2pre-release

Documents are written calmly to the end, and Linux updates by itself again.

Fixed

A long document no longer rushes at the end or freezes the panel. Ask for a PDF or any document and you see it written at an even, readable pace, about 110 letters a second, from the first letter to the last, with any AI model: one that writes slowly, one that writes very fast, or one that sends the whole document at once. Before, the end of a long document came in a blink and the panel could stutter.
The writing cursor stays on the last letter you can see, instead of jumping ahead to where the AI had got to.
Linux updates by itself again. If you have 2026.10.25 or older, it updates like on Windows and Mac. The tray icon next to the clock uses libayatana-appindicator3-1: most desktops have it already. Without it, Cinderpaw still works, and closing the window quits it instead of keeping it in the tray.

New

What's new shows this update's biggest changes, once, the first time you open Cinderpaw after updating: tasks for later, page watches, asking before it pays, and bringing your things over from OpenClaw or Hermes.
v2026.10.26-beta.2 on GitHub →

Cinderpaw v2026.10.26-beta.1pre-release

Documents are written calmly to the end, and Linux updates by itself again.

Fixed

A long document no longer rushes at the end or freezes the panel. Ask for a PDF or any document and you see it written at an even, readable pace, about 110 letters a second, from the first letter to the last, with any AI model: one that writes slowly, one that writes very fast, or one that sends the whole document at once. Before, the end of a long document came in a blink and the panel could stutter.
The writing cursor stays on the last letter you can see, instead of jumping ahead to where the AI had got to.
Linux updates by itself again. If you have 2026.10.25 or older, it updates like on Windows and Mac. The tray icon next to the clock uses libayatana-appindicator3-1: most desktops have it already. Without it, Cinderpaw still works, and closing the window quits it instead of keeping it in the tray.

New

What's new shows this update's biggest changes, once, the first time you open Cinderpaw after updating: tasks for later, page watches, asking before it pays, and bringing your things over from OpenClaw or Hermes.
v2026.10.26-beta.1 on GitHub →

Cinderpaw v2026.10.25

It can do things later, and it still asks first. On Windows and Mac, if you have 2026.10.19, it updates by itself.

Linux: update by hand this time. The tray icon now needs one more system library (libayatana-appindicator3-1). The automatic update cannot install it and would stop halfway, so on Linux this version does not update by itself. Download the file for your system below and install it with your package manager, which brings the library too:

Ubuntu, Debian: sudo apt install ./Cinderpaw_2026.10.25_amd64.deb
Fedora: sudo dnf install ./Cinderpaw-2026.10.25-1.x86_64.rpm

New

Ask it to do something later. Say "remind me in 3 minutes to stretch". Cinderpaw shows a card with the time, and nothing is scheduled until you press Create task. Reminders are free: they use no AI. The new Tasks page lists your tasks, where you can pause, change or delete them and see what each one did.
It can watch a page for you. Ask it to tell you when a page says something, like a price or "in stock". It reads the page itself and tells you only when the words are really there. A watch stops after 30 days unless you renew it, and it costs nothing.
Each task has a spending limit you can change, $5 to start. Reminders and watches cost $0.
Closing the window keeps your tasks running. Cinderpaw stays in the tray, next to the clock, so reminders still arrive. To stop everything, choose Quit from the tray icon. On a Mac, click it in the Dock to open the window again.
When a task is done, a small note in the corner tells you, and it stays until you press Got it.
Picking up where you left off. When a chat ended with work unfinished, Cinderpaw may suggest going back to it: at most one card every 4 hours and three a day. It never starts the work again by itself.
Ask for a PDF and you get a formatted document that downloads as a PDF. Tables come out as tables.
The home screen gives each choice five everyday jobs to start from.
You choose what it may do. Under the box where you type, pick Look only, Can edit files or Full access. Look only reads files and the web and changes nothing on your computer. Full access works in every folder, on every drive, and still asks before it deletes something it cannot undo. The folder with your keys stays closed to it, and commands known to wipe a disk are refused.
It asks before it pays. Cinderpaw can fill a cart for you in its browser, but pressing the button that pays (Place order, Pay, Buy now) on any site shows you a card first, and nothing is bought until you press Yes, pay. If nobody is there to answer, it does not pay.
Coming from OpenClaw or Hermes Agent? During setup, Cinderpaw offers to bring your things over: personality, skills, memory, your AI keys (only if you say yes), your chat apps, your scheduled tasks and your old conversations, which you can read in the sidebar. Chat apps and tasks arrive turned off, so the old agent and Cinderpaw don't both answer. Stop the old one, then turn them on in Connectors and Tasks.
It asks before it changes how it practises. After its Dreams, Cinderpaw shows the one change it wants to make, before and after, and nothing changes until you say yes. Dreams stay off until you turn them on in Settings → Learning.
Connect a service with its link and key. In Capabilities → Add your own, paste the link a service gives for AI assistants and its key, then press Connect. Nothing else needs to be installed. Or ask in the chat: "connect Publora, here is the link", and Cinderpaw connects it after you press Connect on a card. When it needs a key, a password field appears in the chat: the key is saved on your computer and never goes to the AI.
Notion, Linear and other services that use a sign-in connect too. Paste the link and press Connect: your browser opens their sign-in page, you press Allow, and they connect. Nothing to set up on their side, nothing else to install. Cinderpaw keeps the sign-in fresh by itself. The Linear, Jira & Confluence and Intercom cards in Capabilities connect this way now too, so they no longer need Node.js on your computer.
"Something you should know." When Cinderpaw finds something important while it works, a small bar above the box where you type tells you, and it keeps working. Press I already knew this and it flags less of the obvious.
One notice at a time. What's new, the update card and the usage question now take turns, in the same card, in the same top-right corner.
You see when a service is connected. A card shows its logo next to Cinderpaw's and what it can now do, whether you connected it in Capabilities or in the chat.
Cubby has 16 new drawings. It looks at you, turns its head, thinks with a paw on its chin, reads, searches with a lens, writes and talks on the phone, and it changes drawings so it does not repeat the same one. It wags its tail in every one, its paw writes and talks, and its eyes follow the lines it reads. While you type, it types along at its laptop; while you record a voice message, it leans in and listens.
What's new shows a picture of each change.
Kimi has two choices: Kimi Code (subscription) and Kimi API (Moonshot, pay per use), each with the right place to get its key.

Fixed

A request for later waits for you. When you give a time or say "later", Cinderpaw offers a card instead of doing it right away.
It answers in the language you just wrote in, even when it remembers you in another one.
Claude on your own Anthropic key starts on the first try and sends less with every message, because the tool list goes out once and is cached.
Documents are written in order, at a pace you can read, to the end.
After a restart, the model picker names the model in use, and says Starting while it loads.
The window can be moved while it loads and during the first setup.
Your first answer comes sooner. The memory download (630 MB) waits until your first reply has finished.
Chats stay where you left them. Switching chats quickly no longer shows the wrong one, and messages you queued are still there when you come back.
Long writing keeps up. Documents appear in parts as they are written, and the panel no longer falls behind.
A narrow window shows the chat or the side panel, not both squeezed.
The command card shows the result and says why a command failed. Its prompt looks like your own computer's: Windows, Mac or Linux.
Leaving a fullscreen video is faster: about 0.8 seconds instead of 2.2 on Windows. On a Mac the video now fills the screen.
Models: Hugging Face cards stay in place when one opens, and the Downloads and compatibility popups are no longer see-through.
Deleting a model Windows keeps locked works now. Cinderpaw deletes it the next time it starts, before any model loads.
Clicks on sites like Reddit work. When Cinderpaw clicks a link on a page that changes without reloading, it now lands on the new page instead of going back to the old one.
A step where Cinderpaw used an app says what it did, not "1 file".
Security: updated parts with known security problems (zip, webpki, follow-redirects, vite, DOMPurify).
v2026.10.25 on GitHub →

Cinderpaw v2026.10.23-beta.1pre-release

Calmer, and it asks first. If you have 2026.10.19, it updates by itself.

New

You choose what it may do. Under the box where you type, pick Look only, Can edit files or Full access. Look only reads files and the web and changes nothing on your computer. Full access works in every folder, on every drive, and still asks before it deletes something it cannot undo. The folder with your keys stays closed to it, and commands known to wipe a disk are refused.
It asks before it changes how it practises. After its Dreams, Cinderpaw shows the one change it wants to make, before and after, and nothing changes until you say yes. Dreams stay off until you turn them on in Settings → Learning.
Connect a service with its link and key. In Capabilities → Add your own, paste the link a service gives for AI assistants and its key, then press Connect. Nothing else needs to be installed. Or ask in the chat: "connect Publora, here is the link", and Cinderpaw connects it after you press Connect on a card. When it needs a key, a password field appears in the chat: the key is saved on your computer and never goes to the AI.
Notion, Linear and other services that use a sign-in connect too. Paste the link and press Connect: your browser opens their sign-in page, you press Allow, and they connect. Nothing to set up on their side, nothing else to install. Cinderpaw keeps the sign-in fresh by itself. The Linear, Jira & Confluence and Intercom cards in Capabilities connect this way now too, so they no longer need Node.js on your computer.
"Something you should know." When Cinderpaw finds something important while it works, a small bar above the box where you type tells you, and it keeps working. Press I already knew this and it flags less of the obvious.
One notice at a time. What's new, the update card and the usage question now take turns, in the same card, in the same top-right corner.
You see when a service is connected. A card shows its logo next to Cinderpaw's and what it can now do, whether you connected it in Capabilities or in the chat.
Cubby has 16 new drawings. It looks at you, turns its head, thinks with a paw on its chin, reads, searches with a lens, writes and talks on the phone, and it changes drawings so it does not repeat the same one. It wags its tail in every one, its paw writes and talks, and its eyes follow the lines it reads. While you type, it types along at its laptop; while you record a voice message, it leans in and listens.
What's new shows a picture of each change.
Kimi has two choices: Kimi Code (subscription) and Kimi API (Moonshot, pay per use), each with the right place to get its key.

Fixed

Your first answer comes sooner. The memory download (630 MB) waits until your first reply has finished.
Chats stay where you left them. Switching chats quickly no longer shows the wrong one, and messages you queued are still there when you come back.
Long writing keeps up. Documents appear in parts as they are written, and the panel no longer falls behind.
A narrow window shows the chat or the side panel, not both squeezed.
The command card shows the result and says why a command failed. Its prompt looks like your own computer's: Windows, Mac or Linux.
Leaving a fullscreen video is faster: about 0.8 seconds instead of 2.2 on Windows. On a Mac the video now fills the screen.
Models: Hugging Face cards stay in place when one opens, and the Downloads and compatibility popups are no longer see-through.
Deleting a model Windows keeps locked works now. Cinderpaw deletes it the next time it starts, before any model loads.
Clicks on sites like Reddit work. When Cinderpaw clicks a link on a page that changes without reloading, it now lands on the new page instead of going back to the old one.
A step where Cinderpaw used an app says what it did, not "1 file".
Security: updated parts with known security problems (zip, webpki, follow-redirects, vite, DOMPurify).
v2026.10.23-beta.1 on GitHub →

Cinderpaw v2026.10.19

A small update with fixes. If you have 2026.10.16, it updates by itself.

Fixed

Claude keys work again. An Anthropic key now works with every current Claude model. Before, Cinderpaw sent a setting these models refuse, so the agent never answered. Keys that belong to more than one workspace work too.
Long chats no longer lag. Typing, scrolling and new replies stay smooth in a chat with hundreds of messages. Cubby is now drawn as a picture, so its animation no longer slows down the whole page.
The notes in a new chat are easy to read. They sit under the suggestion cards on a solid background, instead of sliding under them.
What's new stays open until you close it. A click next to the card no longer closes it by accident.

Changed

Approval cards show more before you say yes. A new tool shows its code and the sites it may reach. Running a tool says where it may reach. Before you publish on a social site, the card shows the text Cinderpaw typed on that page.
v2026.10.19 on GitHub →

Cinderpaw v2026.10.16

The last release was v2026.08.11, when the app was still called Feral. If you have it, it updates to this one by itself.
Feral is now Cinderpaw. Your chats, settings, connectors, models and saved API keys come with you. On the first start, ~/.feral is copied to ~/.cinderpaw and the old folder is left where it was. On a Mac, the system may ask once whether Cinderpaw can read a key that Feral saved: say yes. The update removes Feral and leaves one app, Cinderpaw. On Windows it goes in your Start menu (and on your desktop, if Feral was there). On a Mac it stays where Feral was, so Finder may still call it Feral. Everything you had comes with you. The command is now cinderpaw. The full story is in the 2026.08.15 and 2026.08.21 sections of the changelog, which were written but never released on their own.

Added

New since the pre-releases

Cubby, and a new look: a warm, light design, and a small mascot, Cubby. Press Listen on any setup step to hear Cubby read it.
Finished work shows up in the chat: a document, table or chart the agent makes is a card you can open, and boards draw charts, diagrams and photos.
Posting waits for you: on a social network, the agent writes the post and opens the page in its browser. You read it and press Post.
Bring over OpenClaw or Hermes from the first screen: your persona and your skills come with you.
A thumbs-up teaches the agent a way of working (beta). The next similar request gets the steps that worked.
Projects keep their own instructions and files for every chat inside them.
One Google button sends a file to Drive, Docs, Sheets or Gmail.
A new conversation already knows you and what you did recently. The Memory page shows what it knows, and Forget really forgets.
Control your desktop by voice and by the agent on macOS and Linux too.
A small campfire game to play while a long task runs (turn it off in Settings).
Install with one command into one folder, ~/.cinderpaw, then cinderpaw open. The everyday CLI commands use the names OpenClaw users already know.
Install a skill from a link: paste a GitHub link to a skill in the chat. Cinderpaw fetches it, checks it for dangerous instructions, shows you every line, and installs it only after your yes.
The agent can build a small tool of its own when none of its tools fits the job. It shows you the code and runs it only after your yes; a tool that uses the internet names the sites it may reach.
Sounds: soft sounds when you send, when a reply finishes and when a switch flips, on by default at 35%. Settings has the switch and the volume.

Getting started

Sign in with OpenRouter instead of pasting a key: one button, your browser opens, and the key goes straight to the system keychain. On a computer with no model at all, the home screen is this button, not a text box nothing can answer.
On a small machine, setup recommends a cloud model first, and says why.
24 more cloud providers, now in Models → Cloud next to the models on your computer.

The chat

A built-in browser beside the chat, with tabs, an ad blocker (Brave's engine), downloads, and a wide mode. The agent uses it by itself, you can watch it work, and a click of yours pauses it.
Artifacts: documents, PDFs, Word and Excel files and small interactive apps the agent makes, in a panel that keeps them after the conversation ends. You can edit, rename, export, and send them to Google Docs. PDFs can be filled in and signed by you.
Replies show diagrams, equations and a copy button; what the agent did reads as steps; tool results stay inside the reply and survive a reload.
Conversations get a short generated title after the first answer, and a name you choose is kept.
Delete a chat with Undo instead of "Are you sure?".
The app reopens the conversation you left.
Star the chats you come back to. Long conversations get a chapter rail and a scrollbar you can grab.
A long paste becomes a small "Pasted text" card, and a pasted link shows the site's icon and name.
Under a reply, "N memories used" shows what it remembered for that answer, and answers list the sources they used.
When work finishes while you are looking somewhere else, a notification tells you.
Drop a file anywhere on the chat page. On Windows, Send to > Cinderpaw from Explorer attaches it.
Alt+Space brings Cinderpaw up from anywhere, with the cursor in the composer. Ctrl+N starts a new chat even while typing; Ctrl+, opens Settings; Esc closes the side panel.

Voice

Speech-to-speech calls with Gemini Live, OpenAI Realtime, or Jev (TypeSafe), which acts on your computer while you are still talking: open apps and folders, play music on Spotify, copy, paste, take screenshots.
A call can be parked in a small call pill instead of ended; the pill says what is running and for how long.
The call screen says which language is spoken, lets you pick the realtime model, shows a question the agent asks so you can answer it, and plays a soft tone when the line opens and closes.

Teammates and connectors

Cowork: named teammates the agent can create, each with only its own tools, talking in a group chat you can read, and stopping to ask you before a gated action.
Chat with Cinderpaw from Telegram, Discord, Slack, WhatsApp and more of the 21 platforms in the catalog; files travel both ways where the platform allows.

Memory and learning

The memory page lists facts as sentences, and every fact has Forget.
Learned procedures appear in the skills menu, with the evidence behind them.

Look and feel

Appearance -> Background: Glass (your desktop shows through) or Solid.
Cubby reacts to what the agent is doing.

Known problems

Voice calls with Gemini are in beta. A Gemini call can stop on its own partway through, and the agent can cut off the last word of a sentence: both come from Google's side, and other apps report them too. When a call stops, the screen says why, and a call that had already started reconnects by itself. OpenAI Realtime is the recommended voice and the first one offered.

Security

A prompt-injection gate checks tool calls before they run.
When someone new writes to your helper on a chat app, Cinderpaw asks you "Is that you?" before it answers them.
When a connector needs a password or token, the agent asks for it in a secure field, never in the chat.

Fixed

The ones you would notice, from the last weeks before this release:

Sign in with Google happens in your own browser, can be cancelled, and Google can be signed out; an expired login asks again.
Google Sheets loads in the built-in browser.
On Windows, voice calls no longer open black terminal windows.
The window fits screens smaller than 1280×800.
A tool the agent built that uses the internet no longer stops the agent from starting.
A second skill added from a link no longer replaces the first.
On a chat app, a message with only a picture or a file no longer answers "Allow" to a question that was waiting.
Approval boxes always show Allow or Yes on top, and Deny or No below.

Changed

Chat mode is gone. Every message now goes to the agent, which can still just talk when that is all you need. The Agent/Chat switch and the Tools menu under the message box are gone with it. Your conversations stay.
The interface is English only for this release, and the language setting is gone. More than 70 languages are planned for the next one. Talking to Cinderpaw in your own language, by text or voice, works as before.
A daily usage summary, only if you turn it on. It is off unless you say yes (on the last setup screen, once on the home screen, or in Settings, Privacy). It sends counts per model: tasks, how they ended, time, tokens, cost, thumbs up and down. Never your messages, files, links or keys. Each summary is saved on your computer before it leaves, and Settings shows it. Everyone's totals are on the new Leaderboard tab. The full list is in PROMISES.md, promise 4.
Cinderpaw counts installs now, once, and says so before it does. Promise 4 used to read "no automatic analytics or crash-report uploads". It no longer does, and the honest thing is to say that in the same release rather than quietly stop being true. On the last screen of first-run setup there is a ticked box: Cinderpaw sends one message containing its version number and your operating system, and, if you press one of the buttons below it, where you heard about Cinderpaw. No identifier, no timestamp of yours, nothing about you and nothing about what you do. Untick it and nothing is sent. Either way ~/.cinderpaw/.install-counted is written with exactly what happened, so the claim is checkable on your own disk.

Why: the GitHub download count is not an install count. v2026.08.11 showed 519 downloads, of which 449 were the updater fetching latest.json from machines that already had Cinderpaw. Around 60 were real installers. We would rather count honestly once than keep quoting a number we know is wrong.

If you already run Cinderpaw, you are never counted: the box only exists during first-run setup, and an update does not run it. So the number counts people who install from this release onwards, and it is a floor, not a total.

Cinderpaw is now open source under the Apache License 2.0. It was source-available under BSL 1.1, which is not OSI-approved: that meant a company had to run a legal review before anyone could try it, and it shut the project out of grants and foundations that require an approved open source licence. Both of those cost more than the licence was protecting.

What changes for you: nothing you have to do, and nothing you lose. Apache 2.0 is free for everyone, including companies of any size. There is no revenue threshold, no commercial licence to buy, and no four-year wait for a version to convert. You may use, modify, redistribute and sell it, and you may offer it as a hosted service, which BSL forbade. The obligation is that you pass on the LICENSE and NOTICE files with any copy and state what you changed.

Apache 2.0 also carries an explicit patent grant, which BSL did not, so contributors and users are covered against patent claims over the code they contributed or received.

v2026.10.16 on GitHub →

Cinderpaw v2026.10.16-beta.8pre-release

The last release was v2026.08.11, when the app was still called Feral. If you have it, it updates to this one by itself.
Feral is now Cinderpaw. Your chats, settings, connectors, models and saved API keys come with you. On the first start, ~/.feral is copied to ~/.cinderpaw and the old folder is left where it was. On a Mac, the system may ask once whether Cinderpaw can read a key that Feral saved: say yes. On Windows, Feral may still be in your list of apps next to Cinderpaw: once Cinderpaw has opened, you can uninstall Feral and keep everything. The command is now cinderpaw. The full story is in the 2026.08.15 and 2026.08.21 sections of the changelog, which were written but never released on their own.

Added

New since the pre-releases

Cubby, and a new look: a warm, light design, and a small mascot, Cubby. Press Listen on any setup step to hear Cubby read it.
Finished work shows up in the chat: a document, table or chart the agent makes is a card you can open, and boards draw charts, diagrams and photos.
Posting waits for you: on a social network, the agent writes the post and opens the page in its browser. You read it and press Post.
Bring over OpenClaw or Hermes from the first screen: your persona and your skills come with you.
A thumbs-up teaches the agent a way of working (beta). The next similar request gets the steps that worked.
Projects keep their own instructions and files for every chat inside them.
One Google button sends a file to Drive, Docs, Sheets or Gmail.
A new conversation already knows you and what you did recently. The Memory page shows what it knows, and Forget really forgets.
Control your desktop by voice and by the agent on macOS and Linux too.
A small campfire game to play while a long task runs (turn it off in Settings).
Install with one command into one folder, ~/.cinderpaw, then cinderpaw open. The everyday CLI commands use the names OpenClaw users already know.
Install a skill from a link: paste a GitHub link to a skill in the chat. Cinderpaw fetches it, checks it for dangerous instructions, shows you every line, and installs it only after your yes.
The agent can build a small tool of its own when none of its tools fits the job. It shows you the code and runs it only after your yes; a tool that uses the internet names the sites it may reach.
Sounds: soft sounds when you send, when a reply finishes and when a switch flips, on by default at 35%. Settings has the switch and the volume.

Getting started

Sign in with OpenRouter instead of pasting a key: one button, your browser opens, and the key goes straight to the system keychain. On a computer with no model at all, the home screen is this button, not a text box nothing can answer.
On a small machine, setup recommends a cloud model first, and says why.
24 more cloud providers, now in Models → Cloud next to the models on your computer.

The chat

A built-in browser beside the chat, with tabs, an ad blocker (Brave's engine), downloads, and a wide mode. The agent uses it by itself, you can watch it work, and a click of yours pauses it.
Artifacts: documents, PDFs, Word and Excel files and small interactive apps the agent makes, in a panel that keeps them after the conversation ends. You can edit, rename, export, and send them to Google Docs. PDFs can be filled in and signed by you.
Replies show diagrams, equations and a copy button; what the agent did reads as steps; tool results stay inside the reply and survive a reload.
Conversations get a short generated title after the first answer, and a name you choose is kept.
Delete a chat with Undo instead of "Are you sure?".
The app reopens the conversation you left.
Star the chats you come back to. Long conversations get a chapter rail and a scrollbar you can grab.
A long paste becomes a small "Pasted text" card, and a pasted link shows the site's icon and name.
Under a reply, "N memories used" shows what it remembered for that answer, and answers list the sources they used.
When work finishes while you are looking somewhere else, a notification tells you.
Drop a file anywhere on the chat page. On Windows, Send to > Cinderpaw from Explorer attaches it.
Alt+Space brings Cinderpaw up from anywhere, with the cursor in the composer. Ctrl+N starts a new chat even while typing; Ctrl+, opens Settings; Esc closes the side panel.

Voice

Speech-to-speech calls with Gemini Live, OpenAI Realtime, or Jev (TypeSafe), which acts on your computer while you are still talking: open apps and folders, play music on Spotify, copy, paste, take screenshots.
A call can be parked in a small call pill instead of ended; the pill says what is running and for how long.
The call screen says which language is spoken, lets you pick the realtime model, shows a question the agent asks so you can answer it, and plays a soft tone when the line opens and closes.

Teammates and connectors

Cowork: named teammates the agent can create, each with only its own tools, talking in a group chat you can read, and stopping to ask you before a gated action.
Chat with Cinderpaw from Telegram, Discord, Slack, WhatsApp and more of the 21 platforms in the catalog; files travel both ways where the platform allows.

Memory and learning

The memory page lists facts as sentences, and every fact has Forget.
Learned procedures appear in the skills menu, with the evidence behind them.

Look and feel

Appearance -> Background: Glass (your desktop shows through) or Solid.
Cubby reacts to what the agent is doing.

Known problems

Voice calls with Gemini are in beta. A Gemini call can stop on its own partway through, and the agent can cut off the last word of a sentence: both come from Google's side, and other apps report them too. When a call stops, the screen says why, and a call that had already started reconnects by itself. OpenAI Realtime is the recommended voice and the first one offered.

Security

A prompt-injection gate checks tool calls before they run.
When someone new writes to your helper on a chat app, Cinderpaw asks you "Is that you?" before it answers them.
When a connector needs a password or token, the agent asks for it in a secure field, never in the chat.

Fixed

The ones you would notice, from the last weeks before this release:

Sign in with Google happens in your own browser, can be cancelled, and Google can be signed out; an expired login asks again.
Google Sheets loads in the built-in browser.
On Windows, voice calls no longer open black terminal windows.
The window fits screens smaller than 1280×800.
A tool the agent built that uses the internet no longer stops the agent from starting.
A second skill added from a link no longer replaces the first.
On a chat app, a message with only a picture or a file no longer answers "Allow" to a question that was waiting.
Approval boxes always show Allow or Yes on top, and Deny or No below.

Changed

Chat mode is gone. Every message now goes to the agent, which can still just talk when that is all you need. The Agent/Chat switch and the Tools menu under the message box are gone with it. Your conversations stay.
The interface is English only for this release, and the language setting is gone. More than 70 languages are planned for the next one. Talking to Cinderpaw in your own language, by text or voice, works as before.
A daily usage summary, only if you turn it on. It is off unless you say yes (on the last setup screen, once on the home screen, or in Settings, Privacy). It sends counts per model: tasks, how they ended, time, tokens, cost, thumbs up and down. Never your messages, files, links or keys. Each summary is saved on your computer before it leaves, and Settings shows it. Everyone's totals are on the new Leaderboard tab. The full list is in PROMISES.md, promise 4.
Cinderpaw counts installs now, once, and says so before it does. Promise 4 used to read "no automatic analytics or crash-report uploads". It no longer does, and the honest thing is to say that in the same release rather than quietly stop being true. On the last screen of first-run setup there is a ticked box: Cinderpaw sends one message containing its version number and your operating system, and, if you press one of the buttons below it, where you heard about Cinderpaw. No identifier, no timestamp of yours, nothing about you and nothing about what you do. Untick it and nothing is sent. Either way ~/.cinderpaw/.install-counted is written with exactly what happened, so the claim is checkable on your own disk.

Why: the GitHub download count is not an install count. v2026.08.11 showed 519 downloads, of which 449 were the updater fetching latest.json from machines that already had Cinderpaw. Around 60 were real installers. We would rather count honestly once than keep quoting a number we know is wrong.

If you already run Cinderpaw, you are never counted: the box only exists during first-run setup, and an update does not run it. So the number counts people who install from this release onwards, and it is a floor, not a total.

Cinderpaw is now open source under the Apache License 2.0. It was source-available under BSL 1.1, which is not OSI-approved: that meant a company had to run a legal review before anyone could try it, and it shut the project out of grants and foundations that require an approved open source licence. Both of those cost more than the licence was protecting.

What changes for you: nothing you have to do, and nothing you lose. Apache 2.0 is free for everyone, including companies of any size. There is no revenue threshold, no commercial licence to buy, and no four-year wait for a version to convert. You may use, modify, redistribute and sell it, and you may offer it as a hosted service, which BSL forbade. The obligation is that you pass on the LICENSE and NOTICE files with any copy and state what you changed.

Apache 2.0 also carries an explicit patent grant, which BSL did not, so contributors and users are covered against patent claims over the code they contributed or received.

v2026.10.16-beta.8 on GitHub →

Cinderpaw v2026.10.16-beta.7pre-release

The last release was v2026.08.11, when the app was still called Feral. If you have it, it updates to this one by itself.
Feral is now Cinderpaw. Your chats, settings, connectors, models and saved API keys come with you. On the first start, ~/.feral is copied to ~/.cinderpaw and the old folder is left where it was. On a Mac, the system may ask once whether Cinderpaw can read a key that Feral saved: say yes. On Windows, Feral may still be in your list of apps next to Cinderpaw: once Cinderpaw has opened, you can uninstall Feral and keep everything. The command is now cinderpaw. The full story is in the 2026.08.15 and 2026.08.21 sections of the changelog, which were written but never released on their own.

Added

New since the pre-releases

Cubby, and a new look: a warm, light design, and a small mascot, Cubby. Press Listen on any setup step to hear Cubby read it.
Finished work shows up in the chat: a document, table or chart the agent makes is a card you can open, and boards draw charts, diagrams and photos.
Posting waits for you: on a social network, the agent writes the post and opens the page in its browser. You read it and press Post.
Bring over OpenClaw or Hermes from the first screen: your persona and your skills come with you.
A thumbs-up teaches the agent a way of working (beta). The next similar request gets the steps that worked.
Projects keep their own instructions and files for every chat inside them.
One Google button sends a file to Drive, Docs, Sheets or Gmail.
A new conversation already knows you and what you did recently. The Memory page shows what it knows, and Forget really forgets.
Control your desktop by voice and by the agent on macOS and Linux too.
A small campfire game to play while a long task runs (turn it off in Settings).
Install with one command into one folder, ~/.cinderpaw, then cinderpaw open. The everyday CLI commands use the names OpenClaw users already know.

Getting started

Sign in with OpenRouter instead of pasting a key: one button, your browser opens, and the key goes straight to the system keychain. On a computer with no model at all, the home screen is this button, not a text box nothing can answer.
On a small machine, setup recommends a cloud model first, and says why.
24 more providers in Cloud Keys.

The chat

A built-in browser beside the chat, with tabs, an ad blocker (Brave's engine), downloads, and a wide mode. The agent uses it by itself, you can watch it work, and a click of yours pauses it.
Artifacts: documents, PDFs, Word and Excel files and small interactive apps the agent makes, in a panel that keeps them after the conversation ends. You can edit, rename, export, and send them to Google Docs. PDFs can be filled in and signed by you.
Replies show diagrams, equations and a copy button; what the agent did reads as steps; tool results stay inside the reply and survive a reload.
Conversations get a short generated title after the first answer, and a name you choose is kept.
Delete a chat with Undo instead of "Are you sure?".
The app reopens the conversation you left.
Drop a file anywhere on the chat page. On Windows, Send to > Cinderpaw from Explorer attaches it.
Alt+Space brings Cinderpaw up from anywhere, with the cursor in the composer. Ctrl+N starts a new chat even while typing; Ctrl+, opens Settings; Esc closes the side panel.

Voice

Speech-to-speech calls with Gemini Live, OpenAI Realtime, or Jev (TypeSafe), which acts on your computer while you are still talking: open apps and folders, play music on Spotify, copy, paste, take screenshots.
A call can be parked in a small call pill instead of ended; the pill says what is running and for how long.
The call screen says which language is spoken, lets you pick the realtime model, shows a question the agent asks so you can answer it, and plays a soft tone when the line opens and closes.

Teammates and connectors

Cowork: named teammates the agent can create, each with only its own tools, talking in a group chat you can read, and stopping to ask you before a gated action.
Chat with Cinderpaw from Telegram, Discord, Slack, WhatsApp and more of the 21 platforms in the catalog; files travel both ways where the platform allows.

Memory and learning

The memory page lists facts as sentences, and every fact has Forget.
Learned procedures appear in the skills menu, with the evidence behind them.

Look and feel

Appearance -> Background: Glass (your desktop shows through) or Solid.
Cubby reacts to what the agent is doing.

Known problems

Voice calls with Gemini are in beta. A Gemini call can stop on its own partway through, and the agent can cut off the last word of a sentence: both come from Google's side, and other apps report them too. When a call stops, the screen says why, and a call that had already started reconnects by itself. OpenAI Realtime is the recommended voice and the first one offered.

Security

A prompt-injection gate checks tool calls before they run.

Changed

Chat mode is gone. Every message now goes to the agent, which can still just talk when that is all you need. The Agent/Chat switch and the Tools menu under the message box are gone with it. Your conversations stay.
The interface is English only for this release, and the language setting is gone. More than 70 languages are planned for the next one. Talking to Cinderpaw in your own language, by text or voice, works as before.
A daily usage summary, only if you turn it on. It is off unless you say yes (on the last setup screen, once on the home screen, or in Settings, Privacy). It sends counts per model: tasks, how they ended, time, tokens, cost, thumbs up and down. Never your messages, files, links or keys. Each summary is saved on your computer before it leaves, and Settings shows it. Everyone's totals are on the new Leaderboard tab. The full list is in PROMISES.md, promise 4.
Cinderpaw counts installs now, once, and says so before it does. Promise 4 used to read "no automatic analytics or crash-report uploads". It no longer does, and the honest thing is to say that in the same release rather than quietly stop being true. On the last screen of first-run setup there is a ticked box: Cinderpaw sends one message containing its version number and your operating system, and, if you press one of the buttons below it, where you heard about Cinderpaw. No identifier, no timestamp of yours, nothing about you and nothing about what you do. Untick it and nothing is sent. Either way ~/.cinderpaw/.install-counted is written with exactly what happened, so the claim is checkable on your own disk.

Why: the GitHub download count is not an install count. v2026.08.11 showed 519 downloads, of which 449 were the updater fetching latest.json from machines that already had Cinderpaw. Around 60 were real installers. We would rather count honestly once than keep quoting a number we know is wrong.

If you already run Cinderpaw, you are never counted: the box only exists during first-run setup, and an update does not run it. So the number counts people who install from this release onwards, and it is a floor, not a total.

Cinderpaw is now open source under the Apache License 2.0. It was source-available under BSL 1.1, which is not OSI-approved: that meant a company had to run a legal review before anyone could try it, and it shut the project out of grants and foundations that require an approved open source licence. Both of those cost more than the licence was protecting.

What changes for you: nothing you have to do, and nothing you lose. Apache 2.0 is free for everyone, including companies of any size. There is no revenue threshold, no commercial licence to buy, and no four-year wait for a version to convert. You may use, modify, redistribute and sell it, and you may offer it as a hosted service, which BSL forbade. The obligation is that you pass on the LICENSE and NOTICE files with any copy and state what you changed.

Apache 2.0 also carries an explicit patent grant, which BSL did not, so contributors and users are covered against patent claims over the code they contributed or received.

v2026.10.16-beta.7 on GitHub →