Embodied Claude workflow
A Claude Code project with MCP servers and hooks: long-term memory, needs, social context, camera, voice and sensors
- 377 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- With optional modules the agent gets camera and microphone access and can post, reply to and delete posts on X.
- Hooks run on every prompt and heartbeat and can trigger actions without a separate user request.
Install
- Works in
Manual install
git clone https://github.com/lifemate-ai/embodied-claude.git
cd embodied-claude
./scripts/setup.sh --profile core --non-interactiveFor Linux, macOS and WSL2. On Windows run `scripts\setup.cmd` with the same flags instead.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- With optional modules the agent gets camera and microphone access and can post, reply to and delete posts on X.
- Hooks run on every prompt and heartbeat and can trigger actions without a separate user request.
This is third-party code. Review the repository files before installing.
What it does
Embodied Claude turns a Claude Code project folder into a persistent companion that remembers past sessions. The core profile sets up four MCP servers: long-term memory, a needs system, social context with people, relationships and boundaries, and the individual-kernel core. It needs no camera, API key or other hardware. Hooks in `.claude/` build one self-world state on every prompt, heartbeat and tool result. That state shapes memory selection, attention, planning and action gating. The authors call it a phenomenal-consciousness candidate architecture and state plainly that it does not prove consciousness. Extra capabilities are added with `scripts/setup.sh` flags: a USB or Tapo camera with pan and audio, Whisper transcription, speech through VOICEVOX, ElevenLabs or Atlas Cloud, X search and posting, host temperature and time. Selected servers go into `.mcp.json`, and the rest are left out.
Who it is for. Researchers and enthusiasts experimenting with persistent AI companions on Claude Code with memory, a camera and voice.
Good fit when
- You want Claude to remember conversations and agreements across sessions without manual notes.
- There is a Tapo home camera, and the agent should see the room and hear sound.
- The agent should answer by voice through a local VOICEVOX engine.
- You want to test how a shared agent state shapes planning and action choice.
Not a fit when
- You need memory for a coding project without a companion or social scenarios.
- The agent must not get camera, microphone or X posting access.
- You need a ready app without setting up a Python environment.
Example request
Remember that my setup check word is lantern.Limitations
You need Git, uv and Claude Code, and uv installs Python 3.13 itself. Until setup has written `.mcp.json`, the hooks deny every write in a session opened at the repository root. The setup script does not install ffmpeg, VOICEVOX, drivers or vendor apps. A plain `uv sync` removes previously installed optional packages, so sync with the needed `--extra` flags. A USB camera in WSL2 needs device forwarding, and temperature sensors are usually unavailable there.
How to disable. Close the session and delete the repository folder with `.mcp.json` and memory data. Individual servers are removed by deleting their blocks from `.mcp.json`.
FAQ
How do I check that Embodied Claude works?
Start `claude` in the repository folder and run `/mcp`: four core servers should be connected. Then ask it to remember a word and ask for it in a later turn.
Do I need a camera or API keys?
No, the core profile runs without hardware or keys. Camera, voice and X are added with `scripts/setup.sh` flags.
What if a server does not connect?
Run `./scripts/doctor.sh --live`. It separates blocking errors from hardware warnings and prints a fix for each.
Where are keys and settings stored?
In `.mcp.json` at the project root. The file stays out of Git, is written with `0600` permissions on POSIX, and an existing config is not replaced without `--force`.
Does it work on Windows?
Yes, the core and hooks run on Windows, Linux, macOS on Apple Silicon and WSL2. Windows uses `scripts\setup.cmd`.
Discussion
No questions or reviews yet.
Related
WorkflowsA Claude Code workspace folder: hooks with a local Ollama detect the project type, pick skills and remind you to update the docs
A Claude Code plugin that gives the agent a name, a personality and memory across sessions, plus chat channels.
An MCP server that gives an agent memory across sessions, projects and machines: decisions, dead ends, preferences and skills
An environment for Claude Code that keeps context across sessions via hooks, ledgers and handoffs, orchestrates agents and saves tokens
ClawMem gives Claude Code, OpenClaw and Hermes local memory: note search, session hooks and a 33-tool MCP server with no API keys
The Supermemory plugin for Claude Code: persistent memory across sessions with automatic prompt recall, conversation capture and a codebase index
README in short
The README goes from a hardware-free quick start to a table of extra capabilities and platform support. Then come a diagram of the hook loop, a map of repository packages and rules for storing settings safely.