wactorz 0.6.0

Resilient AI agents that run 24/7 — built for physical AI.

Docs | Installation | Architecture | Home Assistant Addon | Issues

Wactorz is a runtime for physical AI: LLM-driven agents that live next to the sensors, machines and spaces they act on — not in a cloud notebook. Agents run as long-lived, supervised actors on the hardware you already have: a Raspberry Pi in the garage, a factory gateway, an old laptop, a VM in your closet. You describe what you want in chat; the planner writes the Python, spawns it on a node, and supervises it. When an agent crashes, only that one restarts — with its state intact — and you can migrate an agent to a different machine without losing it.

Everything rides on MQTT, so anything happening inside the system surfaces as a topic external code can subscribe to. Home Assistant talks to it the same way Discord and Telegram do — one channel among several, alongside a REST API and an MCP server. The LLM provider is configurable (Anthropic, OpenAI, Gemini, NIM) or fully local via Ollama, so the system keeps running with no cloud at all.

How Wactorz is different

Most agent frameworks build a crew that runs a task and exits. Wactorz builds a system that keeps running. Agents are long-lived, supervised actors — they persist their state, restart themselves when they crash, and can move between machines — rather than functions you call inside one script.

It's not a replacement for Home Assistant — it sits alongside it, adding an LLM planner and dynamic agents on top of the home you already automate.

Quick Start

git clone https://github.com/waldiez/wactorz
cd wactorz
pip install -e ".[all]"

# Start the MQTT broker
docker compose up -d mosquitto

# Set your provider, model, and key (or put them in .env)
export LLM_PROVIDER=anthropic   # anthropic | openai | ollama | nim | gemini | fake
export LLM_MODEL=claude-sonnet-4-6
export LLM_API_KEY=your-key-here

python -m wactorz

Dashboard: http://localhost:8888.

[!IMPORTANT] Wactorz binds to 127.0.0.1 by default and its agents execute code. Set API_KEY before exposing it beyond loopback — it warns at startup if you expose it without one. See Security before deploying anywhere shared.

If you'd rather skip the clone, pull the image from Docker Hub. To run without an API key, use Ollama:

ollama pull llama3
python -m wactorz --llm ollama --ollama-model llama3

Windows setup is in docs/windows.md; the full set of deployment options lives in docs/deployment.md.

Example prompts

when a person is detected in my pc camera, open the office light
when the door opens, make reachy wakeup
when the light has been on for too long, send me a discord notification

Architecture

flowchart LR
    User["User<br/>CLI, REST, Discord, Telegram, HA"] --> Main["MainActor<br/>intent routing"]

    Main --> Actuate["OneOffActuatorAgent<br/>direct service calls"]
    Main --> Planner["PlannerAgent<br/>pipeline planning"]
    Main --> HA["HomeAssistantAgent<br/>REST + WebSocket"]
    Main --> Chat["LLM reply<br/>streaming response"]

    Planner --> Dynamic["DynamicAgents<br/>LLM-generated runtime code"]
    Actuate --> Bus["MQTT broker"]
    HA --> Bus
    Dynamic --> Bus

    Bus --> Dashboard["Live dashboard<br/>agents, logs, cost, heartbeats"]
    Bus --> Remote["Remote nodes"]
    Bus --> External["Sensors, services, and IoT systems"]

Interfaces

LLM Configuration

Set these three env vars in .env or export them in your shell:

# Options: anthropic | openai | ollama | nim | gemini | none | fake
#   none = no provider at all; fake = deterministic canned replies, calls nothing
LLM_PROVIDER=anthropic

# Model ID — examples:
#   anthropic  →  claude-sonnet-4-6
#   openai     →  gpt-4o
#   ollama     →  llama3
#   nim        →  meta/llama-3.3-70b-instruct
#   gemini     →  gemini-2.5-flash
LLM_MODEL=claude-sonnet-4-6

# Generic key — used for anthropic / openai / nim / gemini
# For Ollama, set OLLAMA_URL instead (default: http://localhost:11434)
# For OpenAI-compatible endpoints (Groq, Together, vLLM…), set OPENAI_URL to redirect
LLM_API_KEY=your-key-here

# Optional — sampling temperature for every LLM call.
# 0 = deterministic (recommended for device control and classification);
# leave unset/empty to keep each provider's own default.
# Ignored on Claude models from Opus 4.7 onward, which no longer accept it.
LLM_TEMPERATURE=0

At startup Wactorz logs the configuration it resolved, so you can confirm it at a glance:

LLM: anthropic/claude-sonnet-4-6 | temperature=0.0

Per-call-site overrides (hybrid setups)

Optionally, route individual call sites to different models with LLM_OVERRIDES — for example run the cheap, high-frequency calls on a local model and keep the planner on a hosted one:

# <site>=<provider>[:<model>], comma-separated. Unlisted sites use the global provider.
LLM_OVERRIDES="intent=ollama:qwen3:4b,actuator=ollama:llama3,planner=anthropic:claude-sonnet-4-6"

Sites: main (conversation), intent (intent routing), planner (pipeline planning/codegen), actuator (one-off device control), ha (Home Assistant agent), dynamic (the get_llm() shim inside generated agents).

To compare models per call site before choosing an override, run the built-in evaluation harness — it scores each site automatically and reports accuracy, latency and cost:

python -m wactorz.evalharness \
  --models "ollama:qwen3:4b,anthropic:claude-sonnet-4-6" --temperature 0

See docs/evaluation.md for the benchmark format and metrics.

Security

Wactorz is under active development. The perimeter is closed by default; the remaining caveats below are about what an authenticated caller can do.

What protects an install:

• The API and dashboard require a key. Set API_KEY and every route, the WebSocket handshake, the Prometheus scrape, and the login flow are authenticated — constant-time comparison, session cookies that survive a restart, and sign-in throttling.
• The server binds to 127.0.0.1. Reaching it from the network is deliberate: set WACTORZ_BIND_HOST and WACTORZ_EXPOSED_OK=1. Startup warns if it is exposed without a key, or with a guessable one.
• The broker requires credentials. Anonymous MQTT is off, and remote nodes are given credentials rather than connecting openly.
• Origin and Host allow-lists guard the HTTP surface and the WebSocket handshake against cross-site requests and DNS rebinding.

What to still assume:

• Agents execute code. The planner generates and runs Python, and remote nodes run code delivered over MQTT. Anyone holding the API key or the broker credentials can run code on the host and on every connected node — treat both as root-equivalent, the same way you would an Ansible control node.
• Generated code is screened by a best-effort blocklist, not a sandbox.

Deployment rules:

• ✅ Set API_KEY before exposing anything beyond loopback.
• ✅ Prefer the Home Assistant add-on, which keeps the UI behind HA's ingress auth.
• ✅ Keep the broker on a network you control, with credentials set.
• ❌ Do not run it as a multi-user or multi-tenant service — there is one key, not per-user accounts, and no isolation between agents.
• If you reach it remotely, prefer a VPN or an authenticating reverse proxy over a bare port-forward, even with a key set.

More detail in docs/security.md. Found a security issue? Please see SECURITY.md rather than opening a public issue.

Repository Map

Documentation

Contributors

Panagiotis Kasnesis 📆 💻

Lazaros Toumanidis 💻 🎨

Chris 💻 📓

Amalia Contiero 💻 📣

Contributions of any kind are welcome. See CONTRIBUTING.md to get started.

Contributing

Read CONTRIBUTING.md for setup instructions, code style, and the PR process.

License

Apache 2.0. Free to use, modify, and distribute.