Skip to content

Unclaw CLI

unclaw-cli gives your existing agentic tools a voice. After one command, agents like Claude Code or the Antigravity CLI can speak their replies aloud through your companion (voice, lipsync, expression) while you keep working in the terminal.

Terminal window
npx unclaw-cli install

That detects every supported agent on your machine and wires them all. To target specific agents:

Terminal window
npx unclaw-cli install antigravity claude

Other commands:

Command What it does
npx unclaw-cli detect List supported agents found on this machine
npx unclaw-cli status Show what is currently connected
npx unclaw-cli test Speak a test line through the avatar (launches Unclaw if needed)
npx unclaw-cli uninstall [agent...] Remove everything it added
npx unclaw-cli doctor Diagnose a connection that is not working
Agent Connection /unclaw command
Claude Code MCP (user scope) skill
Codex CLI MCP (~/.codex/config.toml) skill
opencode MCP skill + command
OpenClaw MCP (mcp.servers) skill
Gemini CLI MCP (~/.gemini/settings.json) command
Antigravity (CLI + IDE) MCP (~/.gemini/config/mcp_config.json) skill
Cursor MCP tool only
Cline MCP tool only
Roo Code MCP tool only
Windsurf MCP tool only

The CLI is deliberately boring and inspectable:

  1. A small runtime is copied to ~/.unclaw/bin/unclaw-speak.mjs. Run with --mcp it is a standard stdio MCP server exposing three tools: speak, launch_unclaw, and unclaw_status.
  2. Each detected agent gets an unclaw entry in its own MCP configuration, written in that agent’s native format.
  3. Agents with an extension point also get an unclaw skill or command (the same SKILL.md everywhere, following the AgentSkills convention), so typing /unclaw launches the companion in passthrough mode and turns on voicing.

uninstall removes all of it. Nothing runs in the background; the runtime only executes when an agent calls a tool.

Start your agent and type /unclaw (or just ask it to “talk to me through Unclaw”). The agent launches the app in passthrough mode and starts calling speak with short conversational lines; your companion says them aloud with matching expression. The mood and action arguments let the agent color a line (“joyful”, “celebrate”) without any extra setup.

Building your own integration instead of using a supported agent? See Build on Unclaw.