The claw-code crate is a rename stub, so the Rust port has to be built

The claw-code crate on crates.io is a rename notice

The first warning in the quick start is about the obvious install command. Installing the claw-code crate from crates.io puts a deprecated stub in place, one that places `claw-code-deprecated.exe` rather than `claw`, and running it prints only `claw-code has been renamed to agent-code`. The instruction is explicit: do not use `cargo install claw-code`.

The alternative names a different binary and a different crate, `cargo install agent-code`, which installs `agent.exe` on Windows or `agent` on Unix, and still not a binary called agent-code. The naming here is genuinely confusing, and the consequence is concrete. A reader who reaches for the crate whose name matches the repository ends up with an executable that does nothing except print a sentence, and the correct upstream install is a different command producing a differently named binary. The repository itself is build from source only, so neither route is how you get the claw CLI from this tree.

The source build ends in claw doctor, not in a working session

The supported path is a clone and a workspace build:

git clone https://github.com/ultraworkers/claw-code
cd claw-code/rust

cargo build --workspace

Credentials come from an Anthropic API key in the environment, set as `export ANTHROPIC_API_KEY="sk-ant-..."`, and the distinction drawn in the guide is that this is an API key and not a Claude subscription. After that the sequence is a health check, one prompt, then the interactive session:

./target/debug/claw doctor

./target/debug/claw prompt "say hello"

./target/debug/claw

That ordering is the honest part of the project. `claw doctor` is named as the first health check after building, and the last command starts an interactive session.

claw acp serve exits 0 and does nothing yet

Editor integration is the clearest unfinished piece. The status note says the project does not ship an ACP or Zed daemon and no JSON RPC entrypoint yet. Two commands exist anyway: `claw acp` and `claw --acp`, which report the current status, and `claw acp serve`, which is described as a discoverability alias only. It returns status and exits with code 0.

That exit code is the trap. A script, a CI check, or an editor probing for a working ACP endpoint will read `claw acp serve` as a success, because the process exits clean, while no daemon ever bound a port. Real ACP support is tracked separately in ROADMAP.md, and the public JSON contract lives in docs/g011-acp-json-rpc-status-contract.md. So the honest reading of a clean exit is that the feature is absent, not that it started. If you are wiring this into Zed or any ACP client today, the project cannot serve you, and the file you want to watch is the roadmap rather than the binary.

Windows renames the binary to claw.exe and adds a shell

The Windows path differs in two ways that break copy and paste. The binary is `claw.exe`, not `claw`, so a path written for Unix will not resolve, and the guide suggests either using `.\target\debug\claw.exe` or running `cargo run -- prompt "say hello"` to skip the path lookup entirely. Onboarding begins with installing Rust from rustup.rs, closing and reopening the terminal, then confirming the toolchain is on PATH:

cargo --version

The clone and build work in PowerShell, Git Bash, or WSL:

git clone https://github.com/ultraworkers/claw-code
cd claw-code/rust
cargo build --workspace

Credentials are set with PowerShell syntax and the prompt uses the .exe form:

$env:ANTHROPIC_API_KEY = "sk-ant-..."
.\target\debug\claw.exe prompt "say hello"

The README also points to the PowerShell-first Windows install and release quickstart in docs/windows-install-release.md.

Local models are a doc link, and the default key is Anthropic

The default credential path is a hosted provider. There is no offline or local option in the quick start; `ANTHROPIC_API_KEY` is the only variable the sequence sets, and the note insists it is an API key rather than a subscription.

For local models the README points at docs/local-openai-compatible-providers.md, which covers local OpenAI compatible providers and offline skill installs. If you arrived wanting to run this against a model on your own machine, the answer is that the settings exist but are not written in the pages a reader is likely to read first, and the practical consequence is that you have to go find the file before you can even know which environment variable or config key to set.

The compose stack runs RAG on mock embeddings by default

The container workflow is defined in docker-compose.yml, and its defaults are the part to read carefully. Three services exist: qdrant on ports 6333 and 6334 with a named storage volume, plus rag-serve and rag-ingest, both built from ./rust using crates/claw-rag-service/Dockerfile. The serve service runs `serve` against /data/index.sqlite, publishes 8787, binds `CLAW_RAG_HOST` to 0.0.0.0, and points at `http://qdrant:6334` with collection `claw_rag_chunks`. Ingest runs the same image with the `ingest` command and mounts the repository read only at /workspaces/main.

The default that changes results is `CLAW_RAG_MOCK_PROVIDERS: "1"`, with a comment saying to override it in your shell for real providers. Mock embeddings are placeholders rather than real embeddings, so retrieval behavior in the compose setup is not retrieval behavior you can evaluate. Two smaller things: qdrant is pinned to the floating `qdrant/qdrant:latest` tag while nothing else in the file is versioned, and neither rag service declares a healthcheck or a restart policy, so a dependency that starts slowly is your compose run's problem.

The project says it is a museum exhibit in the first screen

The most important thing in this repository is a warning, and it appears before any of the documentation. Claw Code is not the serious production project here, and the description calls it an agent-managed museum exhibit, an artifact swept and labeled by agents and automatically maintained, not something meant to be hand operated like a normal product repo. The harnesses do the planning, execution, verification, labeling, and preservation, and the crabs are described as keeping the tank running.

For actual work the project redirects you to two other repositories, LazyCodex at code-yeongyu/lazycodex and Gajae-Code at Yeachan-Heo/gajae-code, and two Discords, ultraworkers and gajae-code. PHILOSOPHY.md holds the intent and system design framing, and the longer public explanation is a post rather than a document in the tree. The judgment to apply: this is worth reading as an artifact and not worth adopting as infrastructure, and the presence of a working build, a doctor command, and a container stack does not change that framing, it just makes the artifact unusually complete.

A Python tree sits next to the Rust one, and the port is checkpointed

The repository shape explains part of the confusion. `rust/` is the canonical workspace and the claw CLI binary. `USAGE.md` is the task oriented guide for the current product surface. `PARITY.md` tracks the Rust port checkpoint and migration notes, and `ROADMAP.md` holds the active roadmap and cleanup backlog. Alongside them, `src/` and `tests/` are described as a companion Python and reference workspace with audit helpers, explicitly not the primary runtime surface.

So the top level also contains install.sh, a Containerfile, docker-compose.yml, scripts/, assets/, how_to_run.md, concept.md, prd.json, progress.txt, and harness configuration directories including .claude/, .omx/, .port_sessions/, .claw.json, and .claude.json. A reader arriving at this tree should assume a Rust port that was measured against an earlier reference implementation, and should read PARITY.md before assuming any feature in the older layout still behaves the same. The latest GitHub release is v0.1.10 from 2026-11-02, and the last push was also 2026-11-02.

Editorial conclusion

Treat claw-code as something to read and audit rather than something to deploy. If you want to inspect the Rust port of the claw harness, build the workspace, run claw doctor before anything else, and expect to supply your own Anthropic API key. Do not reach for the crates.io package, and do not read a clean exit from claw acp serve as working editor integration, because the ACP daemon is tracked as unfinished. If you want to run real agent work today, the project itself points you at LazyCodex and Gajae-Code instead.

Frequently asked questions

Official sources

1. Official README
2. Project repository
3. README
4. Releases