Library Curator
- Desktop app
- All plans
What it does
An agent that tidies the prompt library, the skills library and the memory of a project and of your account in one go. It finds duplicates across the three (a prompt repeating a skill, a project note copying an account note, two notes on one topic), rules that contradict each other, stale entries (shelved work, dead links, agents, ids or files that no longer exist), skills unrelated to the technologies of the repository, and project skills worth sharing with every project. It fixes the safe things itself and asks you on one card for everything else. It also proposes edits to the skills you actually used, learned from what you typed while they were in play.
Where to find it
- Skills library (dock, group Knowledge) > Clean up: the dialog opens on the Library Curator block, above Remove duplicate copies and Empty this library.
- Project memory (dock, brain icon) > Clean up.
- Prompts library (dock) > Clean up.
The button shows when the library belongs to a project (not on an account-only library).
How to use it
- Click Clean up, then Launch the audit. A temporary agent named Library Curator opens in the project, on your default agent CLI.
- It reads the three libraries (project, client folder zone, account) and the repository manifests, then fixes on its own what changes no rule: clearer descriptions, better tags, repaired or removed dead links.
- It writes its findings in its console, then shows ONE approval card: each contradiction as a "which one is right?" question, then deletions, merges, promotions of a project skill to your account and skill edits (with a before / after diff in the preview), each saying what the entry is attached to. Tick what you want; it applies exactly that and nothing else, and re-attaches the surviving skill to the agents, the project and the triggers that used the old one.
- It ends with a short report: what it fixed, what it applied, what it left and why, and the number of entries per library before and after.
Every change it makes is kept in the library history (prompts, skills, memory): any previous version can be restored from there, and its writes are listed under its name in the Changes tab.
Settings
None.
Agent tools (MCP)
library_audit: the curator's audit package (the method, the project's agents, and what the sessions of the last 30 days say about each skill used). Only for the curator.skills_save/skills_deletewithscope: "global": write or delete an account-level skill (promotion = save to the account, then delete the project copy).prompts_update/prompts_deletewithscope: "global": edit or delete an account prompt.memory_save/memory_deletewith their usualscope(project, folder, global).
Providers
Runs on any agent CLI (it uses the AgentsRoom MCP tools). Learning from sessions reads the transcripts of Claude Code, Codex CLI and Antigravity CLI for the skills an agent loaded during a session (and Grok, Mistral Vibe, oh-my-pi and Pi when the skill's SKILL.md file was read); on every CLI, the skills attached to an agent (its defaults and the project baseline) count as used for its whole session.
Mobile
Not on the mobile app (decision of 2026-11-03). The approval card of a curator started on the desktop can be answered from the phone like any other agent question.
Limits
- Started by hand only: no schedule (a scheduled run would be closed by the idle watcher before anyone answers its card).
- Learning from use covers the 12 most used skills of the last 30 days, about 2,500 tokens of your prompts each, secrets masked on the machine.
- One card holds at most 4 questions of 8 items; when everything does not fit, a second card comes after the first is applied, never more, and the rest is listed in the final report.
- It never proposes deleting an account entry just because it does not fit this project: other projects may use it.
- Skills attached to a single ticket, and prompts cited inside ticket bodies, are not checked before a deletion.
- It never touches the Learning Loop notes (
user-lessons,user-lessons-pitfalls), the repository files or git, and never the agents' notebooks nor the assistant's memory. - From the detached Prompts panel, the curator opens in the main window.
Common questions
- Can it delete my skills without asking? No. Deletions, merges, promotions and skill edits all wait for your tick on its card. Only descriptions, tags and links change on their own.
- Can I undo what it did? Yes, from the history of each entry (prompts, skills, memory).
- Does it read my code? It reads the repository manifests (package.json, go.mod...) to know the stack, checks that files named in notes still exist, and may read the git history and the backlog to tell whether a note is stale. It never edits a file and never writes to git.
- Will removing a skill break my agents? It sees which agents, triggers and the project baseline use each skill, says so on the card, and re-attaches the surviving skill after a merge or a promotion.
- Can it run every week? Not yet: it is started from the Clean up button.
Related
- Skills Library: the skills it audits, and
skills_delete. - Project Memory: the notes it audits.
- Prompt Library: the prompts it audits.
- Learning Loop: Learning Loop, whose notes it never touches.