Basou — AI skill for Claude Code

Basou

A harness for steering AI coding agents.

AI coding agents do the typing now. Basou is the harness you steer them from: a **declarative workspace** you drive each project from, an **orientation layer** that carries your intent across sessions, and a replayable, local-first trail of what the agents actually did — all kept in plain files next to your code.

[](LICENSE) [](https://www.npmjs.com/package/@basou/cli) [](#installation)

Basou is local-first and zero-network: it reads the agent logs already on your machine and writes only into a `.basou/` directory beside your repo. Nothing leaves your machine.

Why Basou?

When an agent writes most of the code, the scarce thing is no longer typing — it's **control**: keeping a human in the saddle, steering each project the same way every time, and not losing the *why* between sessions. The answers to "what does this code do?" already live in the code; a read-only assistant can reconstruct them. What no after-the-fact reader can recover is the live context around a decision — the intent, the road not taken, what you were in the middle of. That has to be captured at the moment, and it has to survive the gap to the next session.

Basou is built around those two needs. It is **not** a dashboard you watch or an audit log you file away — it is the tack you actually hold while you work.

The two foundations

🐎 The saddle — a declarative workspace

Declare each repo once — its visibility, its language, where its agent instructions live — in a single manifest, and `basou project` derives the rest: the capture `source_roots`, the `AGENTS.md` / `CLAUDE.md` / Copilot instruction-file wiring, the `.gitignore` entries that keep private canonicals out of public history, and the combined workspace view. You edit the