unclaw

Give any coding agent a voice: wire the UnClaw 3D avatar's speak capability into Claude Code, Codex, opencode, OpenClaw, Gemini CLI, Cursor, Cline, Roo, Windsurf — with one command.

npx unclaw-cli               # detect installed agents + connect them all
npx unclaw-cli detect        # list what's supported / installed
npx unclaw-cli install codex opencode   # only these
npx unclaw-cli status        # agents connected + is UnClaw live + talkativeness
npx unclaw-cli test          # send a test line to the avatar (verify the chain)
npx unclaw-cli doctor        # diagnose why speaking isn't working
npx unclaw-cli uninstall     # remove

(The installed command is unclaw , npm i -g unclaw-cli then just unclaw ....)

Then start your agent — it gains a speak tool. Launch UnClaw in passthrough mode (/unclaw, or open "unclaw://passthrough") and the avatar voices whatever the agent decides to say aloud.

How it works

The whole capability is one stdio MCP server (unclaw-speak.mjs --mcp, exposing speak + launch_unclaw). Because MCP is the common substrate across every modern coding agent, "support a new agent" just means registering that server + dropping short guidance — no per-agent reimplementation. The speak tool's own description carries the usage rules, so agents know how to use it even without an instructions file.

speak(text) → the agent's shim → soul /passthrough/speak → the UnClaw renderer → TTS + lipsync + facial expression on the avatar. The user controls talkativeness + mute inside UnClaw; every speak response echoes those back so the agent self-adjusts.

What it does per agent (all verified against official docs)

| Agent | MCP registration | Guidance | /unclaw (skill / command) | |---|---|---|---| | Claude Code | claude mcp add --scope user | ~/.claude/CLAUDE.md | skill ~/.claude/skills/unclaw | | Codex CLI | ~/.codex/config.toml (startup_timeout_sec) | ~/.codex/AGENTS.md | skill ~/.agents/skills/unclaw (+ legacy /prompts:unclaw) | | opencode | ~/.config/opencode/opencode.json (mcp, type:"local") | ~/.config/opencode/AGENTS.md | skill ~/.config/opencode/skills/unclaw + /unclaw command | | OpenClaw | ~/.openclaw/openclaw.json (mcp.servers) | workspace AGENTS.md | skill ~/.openclaw/skills/unclaw → /unclaw | | Gemini CLI | ~/.gemini/settings.json (mcpServers) | ~/.gemini/GEMINI.md | /unclaw command | | Cursor | ~/.cursor/mcp.json (type:"stdio") | tool description | — (no user commands) | | Cline | VS Code globalStorage cline_mcp_settings.json (autoApprove) | tool description | — | | Roo Code | globalStorage mcp_settings.json (alwaysAllow) | tool description | — | | Windsurf | ~/.codeium/windsurf/mcp_config.json | tool description | — | | Antigravity | ~/.gemini/config/mcp_config.json (mcpServers, CLI + IDE) | tool description | skill ~/.gemini/antigravity-cli/skills/unclaw → /unclaw |

Every agent with an extension point gets a first-class /unclaw that launches passthrough + switches on the speak tool. The rest rely on the launch_unclaw MCP tool (and auto-launch on first speak).

Notes baked in from the research pass:

• SKILL.md is a cross-agent standard (AgentSkills). The same skill file is consumed by Claude Code, Codex, opencode, and OpenClaw — a skill named unclaw even auto-becomes OpenClaw's /unclaw slash command.
• Codex: custom prompts are deprecated in favor of skills, so the skill is primary (the prompt stays as a legacy bonus). We write config.toml directly to set startup_timeout_sec = 30 — Codex's 10s default trips a spurious "failed to connect" on a Node cold start.
• OpenClaw: external stdio MCP-client support is shipped + stable (release 2026.3.31). The server goes under nested mcp.servers (not mcpServers); one entry serves the embedded agent and is projected into delegated coding CLIs.
• The runtime is copied to a stable ~/.unclaw/bin/unclaw-speak.mjs so configs survive npx-cache eviction.
• Every config uses an absolute node path — Dock-launched GUI editors (Cursor/Cline/Windsurf) get a minimal PATH, and in the UnClaw app process.execPath is Electron, not node.
• MCP JSON writes refuse to run if an existing config can't be parsed.
• AGENTS.md is the emerging cross-agent instructions standard (Codex, opencode, OpenClaw, Cursor, Gemini, Windsurf, ...).

Distribution

Two channels, one shared core (lib/installer.mjs):

1. CLI — publish this package to npm; users run npx unclaw.
2. In-app — the UnClaw app's Connect your coding agent screen calls the same detectAgents() / install() core over IPC (one click per agent).

Node.js is required only where the MCP server runs. Gemini CLI users already have it; for the others the installer resolves an absolute node and, if none exists, should prompt to install one.

Cross-platform

macOS, Windows, Linux. Most agents key their config off the home dir (~/.codex, ~/.gemini, ~/.cursor, ...) which is portable; the OS-specific bits , the VS Code globalStorage root (Cline/Roo) and finding node , are handled in lib/platform.mjs (Library/Application Support · %APPDATA% · XDG_CONFIG_HOME; which · where).

Requirements

Node 18+. Node is only needed where the MCP server runs; the installer resolves an absolute node path (Dock/Start-menu-launched GUI editors get a minimal PATH).