v3.21.52026-11-21
Improved
- Changing the natural voice in a voice chat no longer delays the next words: the short sounds it makes ("Okay.", "Mm-hm.") are kept for each voice and prepared when it is not busy.
v3.21.42026-11-21
Safety
- Voice no longer types after something you had already started typing in Claude Code, Qwen Code or Gemini CLI (a half-written message or a slash command), where pressing Return would have sent both. It says so instead, and you send or clear it first.
v3.21.32026-11-21
Fixed
- When your agent spoke while you coughed or made another sound with no words, its line was held back and then never said, and voice stayed on Thinking. The line is now said as soon as you stop.
v3.21.22026-11-20
Fixed
- When Codex asks whether to resume a paused goal, or offers to update itself when it starts, the session now shows that it needs you, and nothing you say by voice is typed into the question.
- Claude Code stopping because its usage credits ran out partway through an answer now shows as limited, not as done.
- Claude Code asking you to sign in again now shows as signed out after its newer messages too: an API key that is turned off, an expired profile login, a Console sign-in it needs, a login another Claude Code window was refreshing, and a Remote Control sign-in that expired.
v3.21.12026-11-20
Improved
- Your agent answers out loud without asking you first. reminal integrate now allows reminal's speak tool, and only that tool, in Claude Code, Gemini CLI and Qwen Code. Run reminal integrate again to get this.
- Qwen Code can answer voice out loud: it now loads reminal's tools when it starts instead of hiding them until it searches for them.
- Voice works with Gemini CLI, Cursor CLI and pi as soon as they start, once set up with reminal integrate: they now tell reminal when they are ready for a message. Run reminal integrate again to get this.
Fixed
- On a device with no voice of its own, voice no longer goes quiet without a word. With Auto it switches to the natural voice; if you picked the device's voice, it says the device has none. Your agent is told a line was not heard instead of being told it was spoken.
- When your agent cuts in while you talk at length, stopping to listen now works. Its "Sorry," plays at once, you get a moment to finish your word, and what you said after it cut in is no longer sent to it, so you hear what it wanted to say instead of an answer to your last few words.
- A goose session asking to allow a tool now shows as needing you, and goose not signed in to its provider shows as signed out.
v3.21.02026-11-19
Added
- Talk to your agent and hear it answer. Tap the mic in a session, speak, and what you say is typed in for the agent; it answers out loud.
- Hearing your voice and making the agent's voice both happen on your own device. Only the text of what you said leaves it.
- Talk at length: what you have said so far reaches the agent every few seconds, so it can follow along, and it can cut in if you asked it to.
- Talk over it to stop it. Hold while you think, so a pause does not end your turn. Mute to talk to someone else in the room.
- Swipe up on the mic to start, stop or pause listening without opening the panel.
- A natural voice on computers, and your device's own voice on phones, or pick either in the settings.
Safety
- Words you say are typed in only when an agent is waiting for a message. They are not typed into a question reminal sees the agent asking on screen, into a plain terminal, or at a password prompt; you are told when that happens.
- Every part of the voice feature is checked against the exact version this reminal expects before it is used.
Good to know
- The first time you start voice it downloads its models once: about 120 MB on a phone, about 210 MB on a computer with the natural voice, and up to 300 MB with the GPU switch on. After that it starts in seconds.
- Your agent needs reminal's speak tool to answer out loud. Agents set up with reminal integrate have it. In Claude Code's manual permission mode, the first answer asks on screen to allow it: choose not to be asked again.
- Tested with Claude Code. With other agents (Qwen Code, Gemini CLI and others), voice may not recognise every question they ask on screen.
- Voice needs a recent desktop or Android browser. Tested on Chrome; Safari, iPhone, Firefox and sessions on Windows are not tested yet.
- A session on a machine with an older reminal says that voice needs a newer reminal there.
v3.20.72026-11-18
Fixed
- A Claude Code session at its fast limit, or out of usage credits, now shows as at its limit with the newest Claude Code wording.
- An Amp session out of included usage now shows as at its limit, whichever plan Amp offers you next.
- An Antigravity CLI session that stopped on a used-up daily quota, a spend cap or prepaid credits gone now shows as at its limit instead of looking finished.
- A Codex session signing in to a gateway, or setting up Amazon Bedrock, now shows as signed out instead of waiting on you.
v3.20.62026-11-17
Fixed
- A session waiting at the Antigravity CLI sign-in code step, or after a sign-in code it could not use, now shows as signed out instead of looking idle or waiting on you.
- A Kiro CLI session that is not logged in now shows as signed out, from its first "Login now?" question through its login methods and device code.
v3.20.52026-11-16
3.20.5 修复了一个输入问题:当智能体向另一个终端输入命令时,即使命令自带换行,也会立即执行。此前这类命令会滞留在输入框里,需要再按一次回车才发出。
v3.20.42026-11-15
3.20.4 修复了「新功能」列表的一个问题:此前它会把 v3.20.3-rc.1 这类测试版本与当前正式版并列显示,甚至提示可升级。现在测试构建不再出现在列表中,也不会被当作升级选项推送给用户,避免误装测试版。