A Rust agent runtime that keeps your keys, data, and GPIO pins local

What it does ZeroClaw is an agent runtime shipped as one Rust binary. It connects to LLM providers including Anthropic, OpenAI, Ollama, and roughly 20 others, ingests messages from over 30 channels—Discord, Telegram, Matrix, email, voice, webhooks, or a local CLI—and executes tools like shell commands, browser automation, HTTP requests, and custom MCP servers. Everything stays on your machine, governed by a single TOML config file.

The interesting bit Most “local AI” projects stop at the chat window; ZeroClaw treats hardware as a first-class citizen through a Peripheral trait that speaks GPIO, I2C, SPI, and USB to Raspberry Pi, STM32, Arduino, and ESP32 boards. It also layers OS-level sandboxes—Landlock, Bubblewrap, Seatbelt, and Docker—so the agent can touch your system without immediately owning it.

Key highlights

• Provider-agnostic with fallback chains: if one provider flakes, the agent routes elsewhere without manual intervention.
• Event-driven SOP engine for cron, MQTT, webhook, or peripheral triggers, with resumable runs and approval gates.
• Cryptographic tool receipts log every action the agent takes.
• ACP channel exposes the agent to your IDE via JSON-RPC over stdio.
• Web dashboard and HTTP/WebSocket gateway for chat, memory inspection, and config editing.

Caveats

• Provider streaming occasionally falls back to non-streaming mode; the README treats this as a known retry path rather than a bug.
• The V3 configuration schema demands at least four specific TOML section headers, and the docs warn that OpenAI Codex auth has subtle rules about where api_key belongs.

Verdict Tinkerers and privacy-conscious developers who want an always-on agent that can read Telegram and toggle a GPIO pin should look here. If you just need a ChatGPT wrapper in your browser, this is overkill.