From install to your first approved patch.
Seven steps, one terminal. At the end you'll have watched Skep sandbox a worker, re-verify its patch on a clean copy, and land it on a review branch only after you said yes.
Install skep
From PyPI. On Linux, install bubblewrap too so the sandbox backend probes clean.
pipx install skep
sudo apt install bubblewrap # Linux only (dnf on Fedora)
Create your profile
Sets up the personal profile in ~/.skep — provider, model, and endpoint for the assistant. Secrets ride by environment-variable name; Skep never stores values.
skep setup --personal
Check readiness
Doctor reports profile readiness, sandbox backend availability, and probes each configured engine binary. Fix anything red before the first run.
skep doctor
Start the daemon
The web UI and HTTP API bind to loopback only, token-gated. The access token prints on startup.
skep serve # then open http://127.0.0.1:8765
Register a repository
In the chat (web or skep chat), make any local directory a first-class workspace. Project setup infers the verify command from the repo's own entry point and pins it — that pin is what re-verification runs.
/workon ~/code/your-app
Pick a coding engine
The built-in worker is the default. To run Claude Code, Codex, or Aider through Skep, set the engine once per project — external engines are forced into sandbox execution and require the pinned verify command.
skep project setup ~/code/your-app --engine claude_code
Run, review, approve
Start a supervised task. When it completes, read the patch and the evidence — including the independent re-verify verdict — then approve. The patch lands on skep/<task_id>; merging into main stays your call, in your normal git flow.
skep run ~/code/your-app "fix the failing test" --execution-mode workspace skep status --personal skep review <task_id> skep review <task_id> --approve
Stuck on a step?
The quickstart has the full detail, including worker adapters and sandbox notes for each platform.