zeroclaw（zeroclaw 搜索）

ZeroClaw

Rust-based autonomous AI assistant infrastructure. Runtime OS for agentic workflows — deploy anywhere, swap anything. Runs on <5MB RAM, <10ms startup, ~8.8MB binary.

• Repo: https://github.com/zeroclaw-labs/zeroclaw
• Docs: https://github.com/zeroclaw-labs/zeroclaw/blob/main/docs/README.md

When to Use

• Setting up ZeroClaw from scratch (install, bootstrap, onboard)
• Authoring or debugging config.toml (providers, channels, memory, security)
• Binding channels (Matrix, Discord, Slack, IRC, webhooks)
• Troubleshooting CLI errors, auth failures, or daemon issues
• Migrating from OpenClaw or other agent frameworks
• Deploying via Docker, native binary, or webhook gateway

When NOT to Use

• Other agent frameworks (LangChain, AutoGPT) — different architecture
• General Rust development unrelated to ZeroClaw

Quick Reference

Installation

# Homebrew
brew install zeroclaw

# Bootstrap (local clone)
git clone https://github.com/zeroclaw-labs/zeroclaw.git && cd zeroclaw && ./bootstrap.sh

# Remote bootstrap
curl -fsSL https://raw.githubusercontent.com/zeroclaw-labs/zeroclaw/main/scripts/bootstrap.sh | bash

# From source
cargo build --release --locked && cargo install --path . --force --locked

# Pre-built binaries: GitHub Releases (Linux x86_64/aarch64/armv7, macOS, Windows)
# Bootstrap flags: --prefer-prebuilt, --prebuilt-only, --docker

Build requirements: 2GB+ RAM+swap, 6GB+ disk. Runtime: <5MB.

Configuration

Edit ~/.zeroclaw/config.toml. Core sections:

[providers]
# OpenAI, Anthropic, OpenRouter, Google, Mistral, xAI, DeepSeek, 70+ more
# Set via env: OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.

[memory]
backend = "sqlite"  # "postgres", "markdown", "lucid", "none"
# Built-in vector search (BLOBs + cosine similarity) + FTS5 keyword search
# No external vector DB needed — hybrid merge of vector + BM25

[runtime]
kind = "native"  # "docker" for sandboxed execution

[channels]
# CLI, Telegram, Discord, Slack, Matrix, Signal, IRC, Email, iMessage,
# WhatsApp, Mattermost, Lark, DingTalk, QQ, Linq, Nostr, Webhook

[security]
# allowlists, pairing, rate_limits, filesystem_scoping

[autonomy]
level = "full"                          # "supervised" (default) or "full"
max_actions_per_hour = 999999           # ⚠️ MUST be > 0 or daemon silently fails
max_cost_per_day_cents = 999999         # same
require_approval_for_medium_risk = false
block_high_risk_commands = false

Architecture

Trait-driven — every subsystem is swappable:

Authentication

OPENAI_API_KEY=sk-... zeroclaw agent                              # env var
zeroclaw auth login --provider openai-codex --device-code         # OAuth device flow
zeroclaw auth paste-token --provider anthropic --profile default  # token paste

Profiles stored encrypted at ~/.zeroclaw/auth-profiles.json (key: ~/.zeroclaw/.secret_key). Manage with auth use, auth status, auth refresh (see Quick Reference).

WARNING (Feb 2026): Claude Code OAuth tokens (Free/Pro/Max) are restricted to Claude Code/Claude.ai only. Do NOT use them via ZeroClaw — violates Anthropic ToS.

Memory System

Built-in full-stack search engine (no external vector DB):

• Vector: Embeddings as BLOBs in SQLite, cosine similarity
• Keyword: FTS5 virtual tables, BM25 scoring
• Hybrid: Weighted merge of vector + keyword results
• Embeddings: OpenAI, custom URL, or noop
• Chunking: Line-based markdown with heading preservation
• Cache: SQLite LRU embedding cache
• Reindex: Atomic FTS5 rebuild + missing vector re-embedding

The agent manages memory automatically via built-in tools.

Common Patterns

Deploy as daemon with Telegram

zeroclaw onboard --interactive          # configure provider + channel
zeroclaw channel bind-telegram 123456   # bind chat ID
zeroclaw daemon                         # start autonomous runtime

Webhook gateway

zeroclaw gateway --port 8080            # or --port 0 for random

Docker sandbox

# config.toml
[runtime]
kind = "docker"

Migration from OpenClaw

zeroclaw migrate openclaw --dry-run     # preview changes
zeroclaw migrate openclaw               # execute migration

Troubleshooting

Common Mistakes

Security

Supports gateway pairing, Docker sandboxing, allowlists (tools/files/channels), rate limiting, filesystem scoping, and encrypted secrets at rest.