ZeroClaw Developer Guide
ZeroClaw is a local-first AI agent runtime written in Rust. You run one binary, connect it to the model providers and channels you want, and keep the agent, its tools, and its data on infrastructure you control. This guide is the practical developer entry point: what ZeroClaw is, how the runtime is structured, how setup and config work, and where to extend it if you want to contribute upstream.
This guide is based on the official ZeroClaw documentation at docs.zeroclawlabs.ai/v0.8.1/en/ and the official zeroclaw-labs/zeroclaw repository, checked on June 26, 2026. The public master branch may move ahead of v0.8.1, so implementation details should always be re-checked against the tag or branch you are actually working with.
In this section​
1. What ZeroClaw is​
The shortest accurate description is:
ZeroClaw is a single-binary agent runtime that connects LLM providers, channels, tools, memory, and a security policy into one local system.
That combination matters:
- It is not just a CLI chat tool.
- It is not tied to one model vendor.
- It is not only for coding.
- It is not a hosted SaaS control plane.
Instead, ZeroClaw is designed around four explicit project opinions:
- You own it
- Security-first, with escape hatches
- Minimal
- Provider-agnostic
Those priorities explain most of the design decisions you see elsewhere in the project.
2. The mental model​
Think of ZeroClaw as five cooperating layers:
| Layer | What it does |
|---|---|
| Providers | Talk to Anthropic, OpenAI, Ollama, and other model backends |
| Channels | Expose the agent through Discord, Telegram, email, webhooks, CLI, ACP, and more |
| Runtime | Runs the agent loop, tool calling, approvals, scheduling, and coordination |
| Tools & memory | Let the agent act on the world and remember useful context |
| Security & config | Define what the agent is allowed to do and how it is wired together |
That is why the project feels broader than tools like Claude Code or Codex CLI. Those are primarily interactive coding agents. ZeroClaw is closer to an agent platform runtime that can power coding workflows, personal assistants, chat bots, automations, and hardware-aware agents from the same core.
3. Why developers would choose it​
ZeroClaw makes the most sense when you want one or more of these properties:
- Local control over agent state, credentials, and execution.
- Multi-channel presence without running separate bots for each platform.
- Provider flexibility instead of committing to one API vendor.
- A strong safety model with approvals, sandboxing, and receipts.
- Extensibility in Rust for channels, providers, tools, and related runtime surfaces.
- A path from personal setup to production-style service mode.
If your only goal is "chat with one model in the terminal," ZeroClaw is usually more platform than you need. If your goal is "run a real agent system I can own and extend," it becomes much more compelling.
4. What makes the docs unusual​
The official docs distinguish between:
- hand-written guides for concepts, setup, architecture, and operations
- generated reference docs for CLI and config
That is an important detail for contributors. If you are looking up a CLI flag or a config key, the generated reference pages are the authoritative source. If you are trying to understand design intent or workflows, the prose guides matter more.
5. Recommended reading order​
Use this order if you are new to the project:
- This page for the high-level map
- Setup and first run to understand the installation flow and runtime surfaces
- Architecture and runtime for the workspace and request lifecycle
- Configuration and security for the real operating model
- Extending and contributing before touching upstream code