CloseClaw is an open-source Python framework for running a local AI agent on your own hardware with Ollama-backed models, local tools, optional voice, and hardware plus simulation primitives for robotics and edge AI workflows.

If you are looking for a robot operating system, a local LLM stack for robots, or an offline-first AI runtime for Raspberry Pi, Jetson, workshop PCs, and secure edge devices, CloseClaw is the hackable local-first path. It is designed for people who want an agent they can inspect, modify, and run without shipping chat history or tool output to a hosted SaaS.

CloseClaw robot booting into genesis on a local workstation

Today, the repository includes:

• A local CLI chat agent backed by Ollama
• Hardware detection with model recommendations for constrained devices
• Built-in local tools for shell, file reads/writes, and file search
• A TOML-based skill system for lightweight local automation
• Conversation history stored locally in SQLite
• A small REST API server for embedding CloseClaw in other systems
• Optional offline voice plumbing with graceful fallbacks when dependencies are missing
• Simulation, body, and brain modules aimed at robotics and embodied-agent experimentation

Most agent frameworks are optimized for cloud APIs first and local hardware second.

CloseClaw starts from a different assumption: if you are running on a Raspberry Pi, Jetson, Mac mini, workshop PC, or a robot on a private network, local control matters. That means:

• local model serving
• local memory
• local tools
• inspectable Python code
• graceful degradation when hardware is missing

Once your dependencies and local models are installed, CloseClaw can run entirely on your own machine.

Tracked CloseClaw rover on a hacker workbench Builder assembling a robot with CloseClaw nearby

git clone https://github.com/BitmapAsset/CloseClaw.git
cd CloseClaw
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -e .
closeclaw init
closeclaw chat

curl -fsSL https://raw.githubusercontent.com/BitmapAsset/CloseClaw/main/install.sh | sh

The installer:

• detects your OS and hardware
• installs the Python package into ~/.closeclaw/venv
• places closeclaw and clawd wrappers in ~/.local/bin
• runs the first boot sequence

Voice support is intentionally optional.

python3 -m pip install -e ".[voice]"

CloseClaw will still run without voice packages. Missing voice components fall back cleanly instead of hard-failing startup.

These commands are implemented in the current codebase:

closeclaw init # detect hardware, create config, help with model setup
closeclaw chat # start an interactive local chat session
closeclaw info # system info + voice support + available models
closeclaw detect # hardware detection summary
closeclaw model list # list local Ollama models
closeclaw model pull gemma4:e4b # pull a model through Ollama
closeclaw skills list # list bundled and user skills
closeclaw skills run hello name=Pi
closeclaw history list # browse local conversation history
closeclaw serve --host 127.0.0.1 --port 8765
closeclaw simulate --hours 2 --environment warehouse
closeclaw body # hardware/body report
closeclaw brain # memory/brain report
closeclaw genesis # first-boot sequence
closeclaw daemon status

CloseClaw is especially useful when you want a local agent adjacent to hardware.

closeclaw detect
closeclaw info
closeclaw body

closeclaw simulate --hours 4 --environment home

The simulation environment is text-based and stdlib-only, which makes it useful for CI, dry runs, and robotics experiments on laptops that do not have GPIO or sensors attached.

closeclaw serve --host 127.0.0.1 --port 8765

This is handy for local dashboards, kiosk apps, workshop tools, or robot control panels that need a simple HTTP interface.

CloseClaw-style robots operating in a modern home

CloseClaw ships with a lightweight skill format built around skill.toml plus a local entrypoint script.

Bundled example skills in this repo include:

• hello
• calculator
• notes
• sysinfo

List them:

closeclaw skills list

Run one directly:

closeclaw skills run calculator expression="(3 * 7) + 2"

User-installed skills live in ~/.closeclaw/skills.

Start the server:

closeclaw serve --host 127.0.0.1 --port 8765

Available endpoints:

• GET /health
• GET /models
• GET /skills
• GET /hardware
• POST /chat

Example:

curl http://127.0.0.1:8765/health
curl -X POST http://127.0.0.1:8765/chat \
 -H 'Content-Type: application/json' \
 -d '{"message":"Summarize the files in this folder","stream":false}'

python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -e ".[dev]"
python3 -m pytest -q
ruff check closeclaw tests

The current tree passes 1,186 tests.

• Contributing guide
• Changelog
• Core architecture notes
• Mesh / embodiment architecture notes

PRs are welcome, especially around:

• hardware detection on more devices
• robotics and embodied-agent integrations
• local skill examples
• install and packaging polish
• docs and reproducible test coverage

See CONTRIBUTING.md.

MIT, see LICENSE.