The open-source, production-grade Rust rewrite of Claude Code.
Drop-in agentic coding assistant. Native tool calling. Multi-provider. Zero compromises.

Quickstart • Why ClawCode • Architecture • Features • Providers • Contributing

# Clone and build
git clone https://github.com/StartripAI/claw_code.git
cd claw_code
cargo build --release
# Set your API key (Anthropic or OpenAI — auto-detected)
export ANTHROPIC_API_KEY="sk-ant-..."
# Launch
./target/release/claw

That's it. ClawCode auto-detects your provider, spins up a session, and drops you into an agentic coding loop with 42 native tools at your fingertips.

100% behavioral parity — verified by an automated compatibility harness scoring against the upstream reference on 9 independent dimensions. Not a toy clone. Not a wrapper. A ground-up rewrite built for production.

┌──────────────────────────────────────────────────┐
│ claw (binary) │
│ CLI • REPL • Server • Voice • Remote • LSP │
├──────────────┬──────────────┬────────────────────┤
│ commands │ hooks │ QueryEngine │
│ 25 slash │ pre/post │ session-aware │
│ commands │ tool hooks │ prompt executor │
├──────────────┴──────────────┴────────────────────┤
│ runtime │
│ AppRuntime • SessionStore • Token Budget │
│ Auto-Compaction • Plugin Events • Config │
├──────────────────────────────────────────────────┤
│ api-client │
│ Anthropic SSE • OpenAI SSE • Mock Provider │
│ Exponential Backoff • Request Timeouts │
├──────────────────────────────────────────────────┤
│ tools │
│ 42 tools: file I/O, search, bash, web, notebook │
│ MCP, tasks, agents, REPL, LSP integration │
├────────┬────────┬────────┬────────┬──────────────┤
│ bridge │plugins │ remote │ server │ voice │
└────────┴────────┴────────┴────────┴──────────────┘

12 workspace crates. Each with a single responsibility. Zero circular dependencies.

• 42 native tools — file read/write/edit, glob, grep, bash, web fetch/search, notebook edit, structured output, agent sub-tasks, and more
• Native JSON tool calling — Anthropic tool_use blocks and OpenAI function calls, not string hacks
• Automatic retries — Exponential backoff (1s→2s→4s→8s→16s) on 429/5xx, configurable max retries
• Request timeouts — Configurable per-provider (default 120s), prevents hanging connections

• Persistent sessions — Resume any previous conversation by ID
• Auto-compaction — When messages exceed threshold, older context is compressed to memory while preserving the system prompt
• Token budget enforcement — Hard cap on total token spend per session, prevents runaway costs
• Session memory — Key facts survive compaction via dedicated memory entries

• Three-tier permission model: read-only → workspace-write → danger-full-access
• Pre/post tool hooks — Intercept, modify, or deny any tool execution
• Pattern-based rules — Glob-style tool permission rules (Bash*, *Tool, exact match)
• Case-insensitive matching — Permission rules work regardless of tool name casing

• Anthropic — Claude 4 / Sonnet / Haiku with streaming SSE
• OpenAI-compatible — GPT-4o, or any API-compatible endpoint (Ollama, vLLM, Together, etc.)
• Mock — Offline development and testing without API keys
• Hot-swap — Switch providers mid-session via /config provider openai

/status /doctor /config /clear /compact /session /permissions /memory /review /export /model /help /bug /init /version /mcp /skills /cost and more.

# Environment variables
export CLAW_PROVIDER=anthropic # or: openai, openai-compatible
export CLAW_MODEL=claude-sonnet-4-20250514 # any supported model
export CLAW_MAX_TOKENS=16384
# Runtime config (persisted to ~/.config/clawcode/config.toml)
/config model gpt-4o
/config permission_mode workspace-write
/config max_turns 50
/config token_budget 500000
/config auto_compact_threshold 200

ClawCode ships with a built-in compatibility harness that scores the implementation against the upstream Claude Code reference:

cargo run -p compat-harness -- score --manifest parity_manifest.toml

{
 "repo1_structural_gate": true,
 "repo1_tool_coverage": 100.0,
 "repo1_command_coverage": 100.0,
 "repo1_ownership_coverage": 100.0,
 "repo2_tool_coverage": 100.0,
 "repo2_command_coverage": 100.0,
 "repo2_real_tool_coverage": 100.0,
 "repo2_moderate_tool_coverage": 100.0,
 "repo2_stub_tool_surface": 100.0
}

9/9 dimensions at 100%. Zero stubs. Zero gaps.

# Run tests
cargo test --workspace
# Strict lint
cargo clippy --all-targets --all-features -- -D warnings
# Dev mode (auto-reload not included, use cargo-watch)
cargo watch -x 'run -- --help'
# Run with mock provider (no API key needed)
CLAW_PROVIDER=mock cargo run

• Ratatui TUI — Full terminal UI with panels, syntax highlighting, and live streaming
• MCP JSON-RPC client — Native Model Context Protocol for external tool servers
• Multi-agent coordination — Parallel sub-agents with shared context
• Persistent transcripts — Searchable conversation history across sessions
• Plugin marketplace — Community-contributed tool packs and hooks

We welcome contributions! Whether it's a bug fix, a new tool, or a performance optimization:

1. Fork the repo
2. Create a feature branch (git checkout -b feat/amazing-feature)
3. Make sure cargo clippy -- -D warnings passes
4. Make sure cargo test --workspace passes
5. Open a PR

MIT — see LICENSE for details.

Built with 🦀 by StartripAI
_{If ClawCode saved you time, consider giving it a ⭐}