Documentation · v0.3.0 · MIT
Disciplined operations for Claude agents.
Keel Skills is a portable governance framework for running Claude agents without breaking things or burning tokens — a three-level authorization model, cost-aware delegation, and file-grounded context discipline. Install it, configure it per project, done.
In plain words
AI assistants that write code can now act on their own — and sometimes they do something you can't undo, like permanently deleting work or publishing to the live system before anyone checked it. Keel Skills is a set of house rules: it lets the assistant handle the small, safe things by itself, but makes it stop and ask you first before anything risky or permanent. You stay in control without watching every step.
The core idea
A four-step test before anything that writes or changes.
Three levels. Only one authorizes execution.
Most "go do it" instructions are L1 or L2
- L1Broad mandate"improve this", "do what's needed", "handle it"Investigate + propose
- L2Mechanism / direction"use a migration", "edit the config", "use a subagent"Still not permission
- L3Explicit scoped approvalapproved action + scope, or a recorded current decisionExecutes — within scope
Read-only & proposals are free · hot / outward / irreversible / structural → L3 · doubt → L3
The full model — hot zones, mechanical propagation, tie-breakers — is on the Authorization model page, and specified runtime-neutral in the spec.
Install
Two lines in Claude Code, then scaffold your policy. Full steps in Getting started.
/plugin marketplace add https://github.com/quitohooded/keel-skills
/plugin install keel-skills@keel-skillsExplore the docs
Install, scaffold an AGENT_POLICY.md, and run your first guarded session.
L1/L2/L3, the four-step test, hot zones, and mechanical propagation.
The single file where your project's hot zones and sources of truth live.
The runtime-neutral spec so Keel can be reimplemented anywhere.