Talk to your agents without leaving your window.
Give each Claude Code agent a name, a voice and its own folder. Say hey and the name from wherever you are, and it comes back when it's done or stuck.
macOS soon. Nothing leaves your machine.
An illustration of the app, drawn for this page.
Three moves. No new window.
Each agent is a real Claude Code session in its own folder. You address it by name, from wherever you are.
- Bit
- Colour
- Voice
- Ava
- Folder
- ~/work/web
# AGENT.mdBit set up with a colour, a voice and a folder
You run the web app tests and
keep the build green.Name an agent
A name, a colour, a voice and a folder. Its role lives in a plain AGENT.md you can read and edit.
One AGENT.md and one WORKLOG.md per agent
hey Block,check the CI pipeline
or holdCtrlSpaceor type itBlockworkingSay hey and the job
Any name works instantly, no training. A push-to-talk hotkey or typing does the same thing.
Custom wake words, no training
- Terminal now
Two files in ci/ changed. Want the diff?
Show meLateror said out loud, in its own voice
It comes back
A native notification, or its own voice. If it has a question, it asks, and a question is never dropped.
Speech queue, native notifications
Demonstrations drawn for this page, not captures.
Built for more than one at a time.
Each feature below comes from the app as it is today, not a roadmap.
Bit
hey Bit,run the tests
hey Loop,draft the release notes
hey Clawd,what is the API returning?
CtrlSpacedoes the same without a wake word#288888
Drawn for this page.
Address one agent from anywhere
Say hey and its name, hold a hotkey, or type. You don't have to switch to its window, or find it.
Wake word · hotkey · typing
hey Bit,run the tests
hey Loop,draft the release notes
hey Clawd,what is the API returning?
CtrlSpacedoes the same without a wake word#288888
It finds you when it's done or stuck
Agents work in the background and report through a native notification or out loud in their own voice.
Notification · spoken reply
Block 2 minStaging is live. Two files in ci/ changed.
Staging is live.
Nothing you didn't agree to
Every tool has three states: ask, always allow, never allow. There is no bypass mode in the UI, including from voice.
Ask · always allow · never allow
Read filesAskAlwaysNeverEdit filesAskAlwaysNeverRun commandsAskAlwaysNeverPush to remoteAskAlwaysNeverIt remembers yesterday
Conversations persist across restarts, and each agent keeps a WORKLOG.md of what it did.
Persistent sessions · WORKLOG.md
WORKLOG.md
## Yesterday
Rewrote the release notes for 0.4.
Open question: keep the beta tag?## Today
Picked up the beta tag questionYou see the cost as it happens
Model, effort and permission mode per agent. A banner shows your rate limit; a meter shows context use.
Quota banner · context meter
Rate limit resets in 2h 10mClawdopushighaskcontext38%
Drawn for this page.
Free. Open source. Your own login.
MIT licensed. It uses the Claude Code subscription you already pay for, and every turn bills that. Running five agents means five sessions on one account, and the app tells you when you hit the limit.
Nothing leaves your machine.
No API keys, no telemetry, no server but Anthropic's. Speech in and out runs locally. Voice input ships off, wake words are a second opt-in, and every tool permission is yours to grant.
- Keys
- Uses the Claude Code login you already have. Nothing to paste.
- Mic
- Off by default. Wake words are a separate opt-in.
- Tools
- Ask, always allow, never allow. No bypass mode.
- Quota
- A banner shows your rate limit as it happens.
Running in a few minutes.
Windows today. macOS is configured but not yet run on a Mac.
Install
Download the Windows installer from the latest release and run it.
Sign in
Use the Claude Code login you already have. No keys.
Name your first agent
Pick a name, a folder and a voice. Say hey and the name.
ClawdBitTerminalBlockLoop
Good questions.
No. It uses the Claude Code login you already have. Each agent is a real Claude Code session, so every turn bills your existing subscription.
Open Room