AutoHarness 🤖 Agent Open source
Self-learning skill layer for Claude Code that distills, merges and prunes skills from your real sessions
- GitHub stars
- 7.2k
- Stars this week
- +2.6k
- Forks
- 562
- Licence
- MIT
- Last push
- 2026-10-02
- Maintainer
- tigerless-labs
/plugin marketplace add tigerless-labs/autoharnessThird-party subagents & agents run with your permissions. Read the source before installing, and prefer pinned versions.
Works with
About AutoHarness
What it does
Autoharness is a Claude Code plugin that writes and maintains skills for you, based on the work you actually do. It watches your sessions through hooks and counts tool calls. Once a session has done enough real work, a background reflection looks at what happened and decides whether a new skill is worth keeping. It can also merge the lesson into an existing skill, patch a skill, or drop it. Over time it archives skills that never get used, so the folder does not fill up with near-duplicates. It only touches skills it created itself; skills you wrote or installed are left alone.
What is inside
- Capture hooks that record each turn, redact it, and trigger reflection after a set number of tool calls
- A reflector that proposes adds, merges and patches but has no write access of its own
- A promoter that checks each proposal and is the only part allowed to write into
.claude/skills/ - A session-start index listing the skills it has written, grouped by category
- A per-skill ledger of why each skill was created or changed, with the evidence behind it
- A
/learncommand to turn the current session into a skill straight away
All timing and size limits are set with AUTOHARNESS_* environment variables.
Works with
Claude Code only. It needs python3 on your PATH and runs on Linux and macOS.
How to install
Run these inside Claude Code, then reload plugins or restart:
/plugin marketplace add tigerless-labs/autoharness
/plugin install autoharness@autoharness
Maintenance and safety
The project is MIT-licensed, actively developed, and uses only the Python standard library. Reflection runs in extra child sessions, so it uses more tokens. The more often it reflects, the more it costs. Captured windows are redacted before they are stored, but the evidence files still hold slices of real transcripts under .claude/, so check what you commit. Uninstalling the plugin leaves its skills and state on disk until you delete them yourself.
Who should use it
It suits Claude Code users who keep solving the same kinds of problems and want those lessons kept as skills without writing each one by hand. If you prefer to curate every skill yourself, or you work on Windows, it is not a good fit.
Pros
- Learns and maintains skills automatically from real work
- Touches only its own skills, leaving yours untouched
- Keeps a ledger and evidence for every change
Cons
- Background reflection spawns extra sessions and uses more tokens
- Claude Code only, needs python3 on Linux or macOS
Similar subagents & agents
All subagent collections →wshobson/agents 🤖 AgentFree
Large plugin marketplace of specialist subagents, skills and commands for Claude Code
Awesome Claude Code Subagents (VoltAgent) 🤖 AgentFree
160+ production-ready Claude Code subagents organised by category
Power Platform Skills 🤖 AgentOpen source
Official Microsoft plugins for building Power Platform apps, pages and flows with Claude Code or Copilot
Awesome Claude Agents 🤖 AgentFree
An orchestrated dev team of Claude Code subagents with a tech-lead coordinator
Contains Studio Agents 🤖 AgentFree
Subagents for a rapid product studio: engineering, design, marketing and ops
Awesome Codex Subagents 🤖 AgentOpen source
175+ Codex-native .toml subagents across 13 development categories