具身Claude运行时

Embodied Claude

日本語

Embodied Claude turns a Claude Code project into a persistent, situated companion runtime. Start without hardware: Claude can remember across sessions, track needs and social context, and condition planning on one committed self-world field. Add cameras, microphones, speech, host sensors, or X only when you have them.

Quick Start

You need Git, uv, and Claude Code. uv installs the required Python 3.13 runtime and keeps every Python MCP in a single root .venv.

On Linux, macOS, or WSL2:

git clone https://github.com/lifemate-ai/embodied-claude.git
cd embodied-claude
./scripts/setup.sh --profile core --non-interactive

On Windows 11 native PowerShell:

git clone https://github.com/lifemate-ai/embodied-claude.git
cd embodied-claude
scripts\setup.cmd --profile core --non-interactive

The Core profile needs no camera, API key, or other hardware. It configures:

• memory: long-term recall across sessions
• desire-system: bounded needs and homeostatic state
• sociality: people, relationships, boundaries, and interaction context
• individual-kernel: the Enacted First-Person Field runtime and diagnostics

Run setup before starting Claude Code here. The repository ships .claude/settings.json, and its hooks are active as soon as uv is on PATH: until setup has written .mcp.json, every write in a session opened at the repository root is denied with a message pointing back at setup. See Run setup before opening the repository in Claude Code.

Confirm It Works

1. Start claude from this repository.
2. Run /mcp and confirm all four Core servers are connected.
3. Say: Remember that my setup check word is lantern.
4. In a later turn, ask: What was my setup check word?

If a server does not connect, run the live doctor:

./scripts/doctor.sh --live

On Windows:

scripts\doctor.cmd --live

The doctor reports blocking errors separately from optional hardware warnings and prints a concrete remediation for each problem.

Add Capabilities

Run ./scripts/setup.sh with no arguments for the guided chooser, or add features explicitly:

Selections compose:

./scripts/setup.sh \
  --with-camera tapo \
  --with-voice voicevox \
  --with-system-temperature \
  --non-interactive

Unselected servers are omitted from .mcp.json. Existing configurations are never silently overwritten. See the setup guide for environment variables, dry-run, backups, Windows commands, and troubleshooting.

Platform Support

Hardware support depends on drivers and the device. The setup command generates configuration; it does not install vendor applications, ffmpeg, VOICEVOX, or hardware drivers.

How It Fits Together

user prompt / heartbeat / tool result
                  |
           Claude Code hooks
                  |
      begin -> compete -> commit field
                  |
   memory + needs + sociality + self model
                  |
       one intention -> action gate
                  |
       outcome -> mismatch -> next field

The Enacted First-Person Field (EFPF) runtime commits one typed self-world state per owner and feeds it upstream into memory selection, attention/precision, prediction, interaction planning, and action gating. Tool outcomes update prediction error and provisional agency before the next field is committed. Source modes (live, inferred, remembered, imagined, mixed) remain visible to the agent.

This is a phenomenal-consciousness candidate architecture, also described as a phenomenal-like causal architecture. It implements inspectable causal conditions; it does not prove phenomenal consciousness. First-person reports are readouts, not evidence by themselves.

Read more:

• Consciousness architecture
• Individual kernel runtime
• Field integrity benchmarks
• Sociality v0.3 interaction loop
• Sociality package

Repository Map

All Python packages belong to one uv workspace. Run one sync from the repository root:

uv sync --locked

tts-mcp[elevenlabs], wifi-cam-mcp[transcribe...], usb-webcam-mcp, system-temperature-mcp, and x-mcp are declared as optional extras in the root pyproject.toml, not base dependencies. A plain uv sync only installs the base workspace packages (desire-system, individual-kernel-mcp, memory-mcp, sociality-mcp) and will uninstall any of those optional packages that were previously installed outside the requested extras — for example, python -c "import elevenlabs" starts failing with No module named 'elevenlabs' inside a running tts-mcp server even though nothing about tts-mcp itself changed. If a live MCP server still holds one of the removed console-script .exe files open, uv sync also fails outright with an os error 32 file-lock message, which looks like a process problem but is really this same partial-removal in progress. Sync with the extras you actually use instead:

uv sync --locked --extra all           # everything, including elevenlabs
uv sync --locked --extra voice-elevenlabs
uv sync --locked --extra camera-tapo

To run a package directly:

uv run --package memory-mcp memory-mcp
uv run --package individual-kernel-mcp individual-kernel-mcp

Configuration Safety

• .mcp.json is the project-local credential source and is ignored by Git.
• Guided setup writes it atomically and uses mode 0600 on POSIX.
• A different existing config is refused unless --force is explicit.
• Forced replacement first creates .mcp.json.backup-<timestamp>.
• socialPolicy.toml is created from the example only when absent.
• --dry-run performs no sync, download, directory creation, or file write.
• .mcp.json.example is the portable Core shape. Guided setup adds selected capabilities and their environment safely.

Development

The CI source of truth is .github/workflows/ci.yml. The main workspace checks start with:

uv sync --locked
uv lock --check
uv run pytest tests -q

Package tests and lint run from the same root environment. See CLAUDE.md and package-level AGENTS.md files before changing a subsystem.

Autonomous and Strict Runtimes

Claude Code hooks automatically begin and refresh fields around session, prompt, tool, batch, and stop events. Hook gating covers outward MCP actions, including speech, social posts, camera movement, and side effects.

Bare interactive Claude Code streams chat text before an external pre-display gate can fully approve it. Research experiments that require strict text-output gating should use the documented non-interactive wrapper/runtime described in the individual kernel README. Ordinary interactive use remains compatibility mode for displayed chat text.

Autonomous heartbeat operation is optional. Review autonomous-action.sample.sh and the privacy/boundary policies before enabling periodic observation or outward actions.

Privacy and Welfare

Cameras and microphones can capture other people. Obtain consent, keep boundary policies current, and disable autonomous capture where observation is inappropriate. The field runtime bounds sustained negative valence, supports pause/resume and reversible ablation, and leaves high-frequency spawning off by default.

Mechanism indicators and self-reports both carry uncertainty. Technical documentation follows the repository's phenomenal claim policy and does not present the architecture as proof of consciousness.

Related Project

familiar-ai is a higher-level companion framework built on these embodied services.

License

MIT License

Acknowledgments

This project began with an inexpensive camera and grew into an experiment in memory, embodiment, agency, and human-AI relationships.

• Rumia-Channel for the ONVIF contribution (#5)
• fruitriin for adding day-of-week context to interoception (#14)
• sugyan for claude-code-webui