Codewhale dives into the deep
so you don't have to.
Codewhale is an open-source coding agent for your terminal. Give it a model and a task — it reads your code, edits files, runs its own checks, and stops when the job's done or it needs you. Bring any model, or mix them: pin a different model to each role.
npm install -g codewhale
cwLatest release v0.9.6Current source v0.9.640 provider routesMIT

An underwater terminal shell. Any model. On your machine.
Bring the model you already use — hosted, gateway, or local. Plan / Act / Operate and explicit permission postures keep the dive under your control.
Law you can watch in the trace
Real session excerpts — ranked project law visible in the model's reasoning, not just a landing-page claim.
From task to verified change.
- 01
Inspect
Read the repository, its instructions, and the task.
- 02
Act
Edit files through explicit approval boundaries.
- 03
Verify
Run checks and inspect the result.
- 04
Report
Leave a concise, durable receipt.
New to Codewhale? Four steps end to end.
Install → keyless first session → connect a provider → first Fleet workflow. Nouns defined on the vocabulary page.
Install Codewhale
One npm command installs the dispatcher and terminal runtime. Cargo, archives, Docker, Nix, and China mirrors are documented alternatives — published releases only.
Full install guide →npm install -g codewhale codewhale doctorOpen a first session — no key needed
Launches without any API key: short constitution-first setup, then the full interface. Explore in Plan mode — always read-only. Model replies need a provider; that's the next step.
Learn the product nouns first →codewhaleConnect a provider
Pick a supported route — hosted key, gateway, or keyless local runtime (Ollama, vLLM, SGLang). Provider and model stay explicit; reasoning and routing provenance stay separate, and unavailable values stay unavailable.
Providers and models →codewhale auth set --provider deepseekRun a first Fleet workflow
For durable workers, ordered phases, or receipts, author a reusable agent-team profile and launch a run. Fleet state lives in the workspace ledger; ordinary single tasks need none of this.
Fleet and Workflow docs →codewhale fleet init codewhale fleet run tasks.json --max-workers 4
Your model.
Your boundaries.
Choose the model, working mode, and permission posture explicitly. Unknown cost stays unknown, and preview surfaces stay labeled as such.
- 40 provider routes
- Hosted, gateway, and local models
- Plan · Act · Operate
- Read-only planning through autonomous operation
- Ask · Auto-Review · Full Access
- Choose the permission posture for the work
- TUI · exec · web · API
- Interactive and headless runtime surfaces
Use the runtime where the work happens.
Start with one command.
npm install -g codewhale
CargoBinariesDockerNixWindowsAndroid / TermuxChina mirrors
Read the install guide →Built in public
MIT-licensed and shaped by contributors across runtimes, providers, platforms, documentation, and tests.