AutoHarness
A self-learning skill layer for Claude Code that distills skills from your sessions, updates them in use and prunes the unused
Medium risk
We rate an entry medium when the tool runs code, makes network calls or reads project files. Check what exactly it does before installing.
Why this level
- Reads the content of your sessions to distill skills from them
- Runs Python hooks, spawns child sessions and writes skill files
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/autoharnessDetects the agents on your machine, checks the risk and pins the version.
Other ways to install
Run one by one in the Claude Code chat
/plugin marketplace add tigerless-labs/autoharness
/plugin install autoharness@autoharnessInstall the plugin: in the Claude Code input box run /plugin marketplace add tigerless-labs/autoharness, then /plugin install autoharness@autoharness and /reload-plugins. python3 must be on PATH.
Other ways from the author
/plugin marketplace add tigerless-labs/autoharness
/plugin install autoharness@autoharnessType the commands in the Claude Code input box. After installing, run /reload-plugins or restart Claude Code. python3 must be on PATH.
This is third-party code. Review the repository files before installing.
What it does
The plugin learns from real Claude Code sessions and turns each episode into a skill without a separate data-collection loop. It folds same-scenario skills into one instead of piling up near-duplicates, updates them as you work and prunes the ones that stop firing. Every session opens with a grouped index of the skills it wrote itself. It touches only its own skills and leaves everything else, yours or installed, untouched. Learning fires automatically after enough tool calls, and the /learn command distills a lesson from the current session on demand. Each create and update is written to a per-skill ledger with its scenario and decision.
Who it is for. For Claude Code users who want their skill layer to build and maintain itself.
Good fit when
- You want good approaches from sessions saved as skills automatically
- You want to lock in a specific lesson with /learn right now
- You want unused skills pruned without manual cleanup
Not a fit when
- You do not work in Claude Code
- There is no python3 on PATH, or the OS is not Linux or macOS
Example request
/learn save the lesson from this session as a skillLimitations
It works only in Claude Code on Linux and macOS and needs python3 on PATH; without it the hooks and the internal MCP server do not run. Third-party marketplaces have auto-update off by default, so updates are manual with a restart. Background reflection spawns child sessions, which spends tokens; the pace is tunable via environment variables. After uninstall the plugin stops running, but the skills it wrote and its state stay on disk and are cleaned up separately.
How to disable. Remove the plugin with claude plugin uninstall autoharness@autoharness and claude plugin marketplace remove autoharness. To clear its data, delete the state directory ~/.claude/autoharness/ or its per-project copy and the plugin's self-authored skills, marked as self-authored, in .claude/skills/.
Security check
- Reads the content of your sessions to distill skills from them
- Runs Python hooks, spawns child sessions and writes skill files
README in short
The README presents autoharness as a self-learning skill layer for Claude Code that distills skills from real sessions, merges same-scenario ones, updates them in use and prunes the unused. The plugin runs entirely on Python with no third-party dependencies, installs through the marketplace with /plugin commands and needs python3 on PATH. Learning fires by tool-call count, the /learn command distills a lesson on demand, and each session opens with an index of its own skills. Configuration is via AUTOHARNESS_* variables, and every decision is logged to a ledger. MIT license.
FAQ
Will it touch my skills?
No. The plugin touches only the skills it created itself, marked as self-authored. Your own and externally installed skills stay untouched.
Do I need to invoke anything manually?
No, learning runs in the background after enough tool calls. The /learn command is only for locking in a lesson from the current session right away.
Related
Open-source personal AI assistant on your own machine: answers in Telegram, Slack, Discord and WhatsApp, extended with skills and plugins
A self-improving agent from Nous Research with a TUI, messaging gateway, cron jobs and skills it writes itself
An open source coding agent for the terminal and desktop with build and plan modes
Open prompt library with a Claude Code plugin, MCP server and CLI: search, fetch and improve prompts and skills from an agent