zeroclaw-labs/zeroclaw

What it is

ZeroClaw is an agent runtime distributed as a single Rust binary. You configure it with one TOML file and run it on your own machine, with your own provider keys and workspace. It connects to LLM providers such as Anthropic, OpenAI, and a local Ollama instance (plus roughly twenty more), receives input over 30+ channels (Discord, Telegram, Matrix, email, webhooks, voice, CLI), and acts through tools including shell, browser, HTTP, hardware peripherals, and custom MCP servers. All inbound messages land in the same agent loop channels overview.

Beyond the chat loop, the binary includes a runtime with separate security, SOP, and memory components architecture overview. Provider fallback chains and routing keep the agent running when one provider fails routing. The default autonomy level is supervised: medium-risk operations need approval and high-risk ones are blocked, with workspace boundaries, command policy, OS-level sandboxes (Landlock, Bubblewrap, Seatbelt, Docker), and cryptographic tool receipts on actions security overview. A Peripheral trait exposes GPIO, I2C, SPI, and USB on Raspberry Pi, STM32, Arduino, and ESP32 hardware. There is also an HTTP/WebSocket gateway with a web dashboard for chat, memory, configuration, cron, and tool inspection, an event-triggered SOP engine, and ACP integration for editors.

It suits people who want a personal agent they host themselves and can extend across messaging platforms and devices, rather than a hosted assistant.

Quick start

On Unix, the recommended path is the generated installer, followed by the one-shot setup wizard Unix install:

curl -fsSL https://raw.githubusercontent.com/zeroclaw-labs/zeroclaw/master/install.sh | sh
"${CARGO_HOME:-$HOME/.cargo}/bin/zeroclaw" quickstart

quickstart asks you to pick a provider and writes a working configuration to ~/.zeroclaw/config.toml. From there the main commands are:

zeroclaw agent -a <alias>         # interactive chat using the [agents.<alias>] entry
zeroclaw service install          # register as systemd/launchctl/Windows Service
zeroclaw service start            # run it always-on in the background

The result is an interactive chat session with your configured agent, optionally turned into a background service. The Windows path does not require Rust; use the prebuilt PowerShell installer in the Windows setup guide, which installs the current release, updates PATH, and launches Quickstart. Source builds and guided installation options are compared in the install section of the quickstart, and the installer's flags are listed by ./install.sh --help.

If you already authenticate with the OpenAI Codex CLI, Quickstart can import that session instead of asking for a key:

zeroclaw auth login --model-provider openai-codex --import ~/.codex/auth.json
zeroclaw quickstart --model-provider openai-codex --model gpt-5.4

A hand-written V3 config needs a provider entry, an agent that references it, and a risk profile, with at minimum four <type>.<alias> section headers and schema_version = 4 provider configuration.

Constraints

• Autonomy is supervised by default; reducing friction means changing the risk profile or enabling YOLO mode, which the docs position for trusted development machines only. The security model is policy-based, so expect approval prompts for medium-risk tool use.
• Configuration is a single TOML file with a versioned schema (currently 4); older or hand-edited files need the documented section structure to load.
• Provider access is required. Subscription auth such as OpenAI Codex or Claude Max setup tokens is stored in auth profiles rather than as api_key fields on the provider entry provider configuration.
• Hardware channel support depends on the Peripheral trait and the listed boards, so it is not a generic "any device" claim.
• Building or syncing the translated docs needs the docs/book/po git submodule; the Rust build does not.
• Dual-licensed MIT OR Apache-2.0; the ZeroClaw name and logo are trademarks of ZeroClaw Labs.

HotRepo's take

The project has attracted attention well beyond a niche runtime: about 33k stars and 5k forks since its February 2026 creation, and a February 17 Hacker News thread. The two Recorded HN threads are thin, though: one announcement with 6 points and 2 comments, and a later thread reporting a 404 on the repository. The notable signal is a skeptical comment in the announcement thread, where a reader said the security discussion focuses on things like gatekeeping who can talk to the bot over Telegram and does not acknowledge problems such as the lethal trifecta discussion. That is one commenter's assessment rather than a verified finding, but it points at the exact area a self-hosted agent runtime should expect scrutiny: the gap between policy-level controls and prompt-driven exfiltration risk. Readers evaluating the supervised-by-default model should treat that distinction as the open question.