lamarck

    1

    A production self-evolving system for agent skills, driven by real invocations.

    Free

    1 useSecurity scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    lamarck

    Example session with this skill installed

    /lamarck audit my-deploy-skill

    • Read your context and instructions
    • Compiled the lamarck

    Summary

    Two separate calls of my-deploy-skill missed the same step. Both ran a deploy command while the git working tree was dirty. The evidence gate is open. No file has been edited yet.

    Evidence

    CheckResult
    Same kind of gap2 separate calls
    Trust tier for this skillevolve (asks you before any edit)
    Default for skills you have not listedobserve (log only, no edits)
    Where the evidence liveslocal ledger only (data/ledger.jsonl)

    Proposed edit

    Insert one condition in that skill's SKILL.md: if git status shows a dirty tree, stop and list the dirty paths before any deploy command.

    The proposal is one insertion. Replay cases harvested from both calls are attached. The background worker will replay them before the edit is kept.

    Next steps

    1. Apply now. The worker edits the skill, replays the cases, and writes one line to data/inbox.md.
    2. Suggestion only. Save suggestions/my-deploy-skill.md and leave the skill file unchanged.
    3. Reject. The reason is stored in data/rejected.md so this proposal is not repeated unless new evidence appears.

    lamarck itself is not on the auto tier. A plugin or a synced skill would only get a suggestion. It would not be edited in place.

    Insert one condition in that skill's SKILL.md

    Connects securely to your tools. The creator never sees your data.

    About this skill

    lamarck is a production self-evolving system for agent skills. Install it once. Every installed skill keeps a record of real invocations, and it changes only when an evidence gate opens.

    It is not a one-shot optimizer for one hand-picked skill. Adapters for Claude Code, Codex, Cursor, and pi write one local ledger. Telemetry stays on your machine. License: MIT. Source: https://github.com/newdee/lamarck-skill

    What it does

    • Logs every skill call and stamps it with that skill file's content hash.
    • Runs a short evaluation at the end of a turn. You choose every turn, manual only, or a batch (default: 5).
    • Turns one user correction into a rubric entry for that skill.
    • Proposes a small edit only after two separate calls show the same kind of gap.
    • Asks you to apply the edit, keep it as a suggestion, or reject it. Auto-apply is opt-in, and a failed replay rolls the edit back. lamarck itself never runs on auto.
    • Prunes instructions that go unused.

    How it works

    1. A hook appends each call to data/pending.jsonl.
    2. A stop hook runs the evaluation. Verdicts go to data/ledger.jsonl. Replay cases come from real traces.
    3. When the gate opens, you get three choices. If you apply, a background job edits the skill, replays the cases, writes the ledger, and notifies you.
    4. If the new version does worse than the old one, lamarck proposes a rollback.

    Install

    npx lamarck-skill
    

    Requires Node.js 18+ and git. The installer wires Claude Code hooks, then runs a self-test. Codex, Cursor, and pi use the adapters shipped in the package.

    Why the gate matters

    Edits stay small, need your approval, and can be reverted. Plugins and synced skills are never edited in place. A skill with a long clean streak drops to spot-checks instead of being rewritten every turn.

    Needs Node.js 18+ and git. No edit is proposed until two separate calls show the same kind of gap, so the first days are often quiet. The default trust tier is observe: evidence is kept, files are not changed. Plugins, marketplace skills, and synced skills are never edited in place. lamarck itself cannot be set to auto. The pi adapter is not yet confirmed in a live pi session.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 1 month ago

    • Free with an Agensi account

    Requires

    Node.js
    lamarck-skill
    github
    macOS
    user

    Permissions

    Terminal / Shell
    Read Files
    Write Files

    File Scopes

    **/SKILL.md
    **/.claude/settings.json

    The hooks run a local Node script. They read the transcript and skill files, and they write the local ledger under data/. A skill file changes only after you approve an edit. The hooks do not call a network API and they do not need a connected account. npm is used only at install time.

    Listed1 month ago

    What's inside

    Frequently Asked Questions

    Use any skill, then check that data/pending.jsonl grew. Unexpected hook errors are written to data/hook-errors.log. An empty error log means the hooks are healthy.

    Claude Code is wired by the installer. Codex 0.142 or newer, Cursor, and pi ship as adapters and share one ledger. The pi adapter matches the published API and has not yet been run in a live pi session.

    Only skills you put on the evolve or auto list. The default tier is observe: calls are logged and nothing is edited. Plugins, marketplace skills, and synced skills are never edited in place.

    Yes, if you promote that skill to auto. The edit must pass replay or it rolls back, and you still get a report. lamarck itself, its iron rules, and plugins always need an explicit yes.

    Run it to drain a backlog, audit one skill, read stats, or read the evolution report. You can also set the trigger to every turn, manual, or a threshold, and set the update cadence to weekly, daily, or continuous.

    Create a file named off in the skill directory to silence the hooks. Manual /lamarck still works. Uninstall with npx lamarck-skill uninstall. That removes the hooks and keeps your files and telemetry.

    New in Agents & Orchestration