# AutoHarness

> A self-learning skill layer for Claude Code: distills skills from your real sessions, updates them as you work and prunes unused ones.

AutoHarness is an open-source Claude Code plugin from Tigerless Labs that maintains a skill library for you. It watches your real sessions, distills reusable lessons into native Agent Skills, folds same-scenario skills together instead of piling up near-duplicates, updates them when you correct them, and archives the ones that stop getting used. It only ever touches the skills it wrote itself.

What it does
- Learns in the background: once a session has done enough work (every 50 tool calls by default), a reflection reads that stretch of the session and proposes new skills or changes to existing ones
- /learn distills the session you're in right now, when you've just worked something out and want it kept
- Patches and merges instead of duplicating: a correction updates the existing skill, and a periodic curator pass folds narrow skills into broader ones
- Opens each session with a grouped index of the skills it wrote, plus a one-line report of what the last run added or rejected
- Keeps a skill by whether later sessions actually use it, not by a benchmark score; unused skills are archived (moved aside, never deleted) and can be restored
- Logs why each skill was created or changed, with the redacted session excerpt that justified it

How it works
- Skills land as plain SKILL.md folders in your project's .claude/skills/ or, for techniques that aren't tied to one repo, in ~/.claude/skills/. Its own state lives in .claude/autoharness/.
- Reflection runs as a separate background Claude Code session on Haiku, with read-only tools and one staging tool, so it can only propose. A deterministic validator checks each proposal (safety, structure, size, self-authored only) and is the only part that writes.
- By default the reflector gets a redacted window of the session plus a short digest of what came before.
- It's tunable with AUTOHARNESS_* environment variables: how often it learns and consolidates, a project-only mode, the session-start index, how many skills each layer keeps, and optional desktop or custom notifications.

How to install it
- It needs Python 3.11 or newer as python3 on your PATH, with no third-party dependencies.
- In Claude Code: /plugin marketplace add tigerless-labs/autoharness, then /plugin install autoharness@autoharness, then /reload-plugins.

Good to know
- It's MIT-licensed, from Tigerless Labs, and built for Claude Code.
- Background reflections are extra Claude Code sessions, so they count toward your usage. They run unattended with permission prompts skipped, limited to read tools and the staging tool.
- Uninstalling stops it, but the skills it wrote and its state stay on disk until you delete them. Your own skills are never touched.

- Type: Plugin
- Made by: tigerless-labs
- Category: AI & agents
- Formats: Claude plugins
- Version: latest
- Website: https://github.com/tigerless-labs/autoharness
- Source: https://github.com/tigerless-labs/autoharness
- Page: https://pluginboard.ai/plugins/autoharness

## Works with

- Claude
- GitHub Copilot
- Claude Code
- OpenClaw

## Install

### Claude

In Claude (web or desktop): Customize → Plugins → Add → Add marketplace, enter tigerless-labs/autoharness, then add autoharness

### Claude Code

In Claude Code:

```text
/plugin marketplace add tigerless-labs/autoharness
/plugin install autoharness@autoharness
```

---
From Pluginboard (https://pluginboard.ai).