claw-bot

Agent Harness for robots

claw-bot

A plugin layer that lets decision agents use robot perception, action, feedback, realtime voice, and field acceptance tools across realworld and simulator environments. 

claw-bot plugin + decision agent = Robot Brain
Install plugin Configure MCP Read GitHub docs
claw-bot dashboard showing environment, perception, action, and MCP trace panels
Dashboard preview for environment status, perception, action payloads, and feedback.

One plugin surface

Realworld and simulator use the same robot interface.

claw-bot keeps environment differences in configuration and adapter capabilities. Agents call the same tools, inspect the active environment, and route actions through structured feedback.

Perception

Snapshots and stream handles

Expose camera, depth, and other modalities as metadata-rich MCP responses without flooding normal tool calls.

Action

Structured dispatch

Compile commands into environment-scoped action payloads for move, turn, navigate, behavior, and stop requests.

Feedback

Plan from controller state

Record lower-controller feedback so the decision agent can continue, recover, or stop with evidence.

MCP surface

Six tools, stable names.

The MCP server exposes a small, predictable robot surface. Switching between realworld and simulator changes capabilities, not tool names.

robot_status robot_configure_environment robot_perceive robot_act robot_subscribe robot_feedback

Field acceptance loop

Voice command to visible robot result.

Live acceptance is an evidence chain, not just a green service check. claw-bot records each handoff from speech through controller feedback.

  1. 01Spoken command

    Operator gives the robot task.

  2. 02Transcript

    Realtime voice or transcribe bridge captures text.

  3. 03Agent plan

    Decision agent uses the claw_bot skill and MCP tools.

  4. 04Action payload

    claw-bot emits structured environment-scoped commands.

  5. 05Controller feedback

    Lower controller reports accepted, blocked, or completed state.

  6. 06Visible result

    The robot or simulator outcome matches the acceptance note.

Install paths

Use npm, GitHub, or MCP registry launch.

The npm package is the canonical distribution. GitHub package install supports source-based demos, and MCP clients can launch the stdio server directly.

npm install and configure

npm exec --yes --package @bozhi-ai/claw-bot@latest -- claw-bot install --yes --json

GitHub source distribution

CLAW_BOT_INSTALL_PACKAGE_SPEC=github:bozhi-ai/claw-bot \
npm exec --yes --package github:bozhi-ai/claw-bot -- claw-bot install --yes --json

MCP stdio launch

npx -y @bozhi-ai/claw-bot@latest mcp

Product boundary

claw-bot is the plugin layer, not the robot runtime.

claw-bot owns

  • claw_bot skill and MCP server
  • CLI and realtime voice control plane
  • dashboard and field acceptance tooling
  • benchmark harness and package metadata

Environment owns

  • sensor setup and actuator firmware
  • low-level safety loops and controllers
  • trained skills, policy networks, and simulator hosting
  • vendor deployment lifecycle and robot model internals