What is Claw Code?
Claw Code is a terminal coding agent, the `claw` CLI, written in Rust. Sigrid Jin started it on 31 March 2026, hours after Anthropic shipped Claude Code 2.1.88 to npm with a source map that exposed 512,000 lines of TypeScript. It began as a Python rewrite of the harness's architecture and moved to Rust within days.
The project now describes itself as an agent-managed exhibit rather than a product: coding agents plan, write, review and release the code, coordinated through Discord, and the README sends anyone who wants to get work done to two other harnesses, LazyCodex and Gajae-Code. It builds from source only, needs an API key (Claude subscription login is not supported) and is not affiliated with Anthropic.
Is Claw Code worth it?
The star count measures a news event, not the software. Claw Code passed 100,000 stars in its first 24 hours and had about 136,000 by 2 April, when the Claude Code leak was the story of the week. It now shows 194,781 stars and 107,303 forks against 1,876 commits. Forks on that scale are mostly people taking copies in case of a takedown, not contributors. Treat the number as a measure of curiosity about the leak.
The code itself is a real Rust workspace. The parity notes, last updated on 3 April 2026, list 9 crates and 48,599 lines of Rust written in four days, with a mock Anthropic service that replays 12 scripted scenarios covering tools, bash permissions, plugins and session compaction. Releases continue: v0.1.13 shipped on 9 November 2026, and recent commits add retries for overloaded responses and setup notes for OpenAI-compatible gateways. Local models through Ollama, llama.cpp or vLLM are documented.
The catch: the maintainers have stepped back from it as a product. The README calls it a museum exhibit kept alive by agents and opens with links to LazyCodex and Gajae-Code instead. There are no published binaries on crates.io (the `claw-code` crate there is a deprecated stub that installs something else), no ACP or Zed support yet, and the legal footing remains awkward for a project that exists because of a leaked codebase, even though independent audits found no Anthropic code in it.
Our take: it is worth an afternoon for anyone studying how agent harnesses are built or how an agent-run repository behaves. As a daily coding agent, OpenCode, Codex and Pi are maintained by people who want you to use them.
Where it earns its place
A platform engineer designing an internal coding agent
See how tool calls, permissions and session compaction are wired in a working harness
Read the runtime crate and run the mock parity harness, which replays scripted Anthropic responses without spending tokens. The only cost is build time: a release build takes 5 to 10 minutes according to the README.
A researcher writing about the Claude Code leak and its aftermath
Document what the community built from the leaked architecture
The repository history, parity notes and philosophy document record how agents rebuilt the harness in Python and then Rust within days, with commit-level evidence. Nothing to buy; the API key is only needed to run prompts.
Before you commit
- Setup
- Install Rust, clone the repository, run `cargo build --workspace` in `rust/`, export an Anthropic or OpenAI API key and run `claw doctor`. Do not use `cargo install claw-code`.
- Integrations
- Anthropic and OpenAI APIs, OpenAI-compatible gateways and local servers such as Ollama, llama.cpp and vLLM; MCP and LSP clients; plugins and skills.
- Data and privacy
- Prompts and file contents go to whichever provider you configure. The project is not affiliated with Anthropic and holds your provider key in the environment.
- Limits
- Build from source only; Claude Pro or Max login is not supported; no ACP or Zed daemon; maintenance is done mainly by agents, so expect churn rather than a stable roadmap.
What’s good
- Readable Rust implementation of a full agent harness
- Mock parity harness with 12 scripted scenarios
- Works with Anthropic, OpenAI-compatible and local models
- Still shipping releases in November 2026
What’s not
- README says it is not the serious production project
- Star count reflects the leak, not adoption
- Build from source only; the crates.io package is a stub
- No Claude subscription login and no ACP support
Best for
Engineers who want to read how a Claude Code-style harness fits together, not to adopt one.
Claw Code alternatives
Claw Code: common questions
- Is Claw Code worth it?
- Claw Code is a piece of GitHub history with a working CLI attached, and its own maintainers point you elsewhere for real work. We rate it 2.5 out of 5.
- How much does Claw Code cost?
- Free and open source (MIT). You pay the model provider whose API key you configure.
- What are the downsides of Claw Code?
- README says it is not the serious production project; star count reflects the leak, not adoption; build from source only; the crates.io package is a stub; no Claude subscription login and no ACP support.
- Who is Claw Code best for?
- Engineers who want to read how a Claude Code-style harness fits together, not to adopt one.
- Is Claw Code safe for company data?
- Prompts and file contents go to whichever provider you configure. The project is not affiliated with Anthropic and holds your provider key in the environment.
- What are the best Claw Code alternatives?
- The closest alternatives we have reviewed are OpenCode, OpenAI Codex CLI and Pi.
Answers checked Nov 14, 2026.