Curator

# curator <p align="center"> <img src="docs/media/demo.gif" alt="curator dashboard demo" width="760" /> </p> > Open-source local dashboard for your Claude Code skills and MCP servers. Free, no login, no API key. See your library at a glance, find duplicates, decide what to keep. ## What it does `curator` runs a local web dashboard at `http://127.0.0.1:4711` so you can actually see what's in your skill and MCP setup. - **At-a-glance overview** — total skills, always-loaded vs on-use token cost, grade distribution, top-heaviest descriptions, top-heaviest bodies, MCP servers. - **Cluster overlap analysis** — silhouette-scored hierarchical clustering surfaces overlapping skills. Click any cluster to see shared core vocabulary, per-skill unique vocabulary, and pairwise similarity. Members with no unique vocabulary are flagged as likely duplicates. - **Topic browsing** — auto-detected tag groupings (`kotlin`, `testing`, `frontend`, …) in the sidebar. A skill can belong to several. - **Edit and delete in place** — click a user-owned skill, edit the body, save. Or delete it. Plugin-installed skills are read-only. - **Backups before every change** — written to `~/.claude/curator/backups/`. Every edit and delete is reversible. Nothing auto-applies. You decide what to keep. By default, skill scans include: - `~/.claude/skills` - `~/.claude/plugins` - `~/.agents/skills` - `.agents/skills` in the current workspace and its parents To scan additional workspace-local skill libraries, pass a path list with your shell's path separator: ```bash CURATOR_EXTRA_SKILLS_DIRS=/path/to/project/.agents/skills curator scan ``` ## Install ```bash git clone git@github.com:KarimJebara/Curator.git cd Curator npm install npm link # puts `curator` on PATH; one-time setup ``` `npm link` is optional — you can also run everything as `node bin/curator.js <cmd>`. ### Optional: install as a clickable Mac app If you want a `Curator.app` you can launch from Spotlight or the Dock instead of running `curator dashboard` from a terminal: ```bash bash scripts/build-app.sh cp -r dist/Curator.app /Applications/ ``` Clicking the icon opens a Terminal window that runs `curator dashboard`. Closing the Terminal window stops the server. Requires `npm link` first so `curator` is on PATH. ## Quickstart ```bash curator scan # deterministic snapshot, ~3 seconds, no API key curator dashboard # opens http://127.0.0.1:4711 in your browser ``` That's it. The default scan is fully deterministic and finishes in seconds. There's also a weekly cron option if you want it to re-snapshot on its own: ```bash curator schedule install ``` ## How it works A single `curator scan` produces two views of your library: **Clusters** — Ward-linkage hierarchical clustering on TF-IDF + bigrams, with silhouette confidence scoring. Each cluster gets a high (≥0.28), medium (0.18–0.28), or low (<0.18) confidence badge so you know how much to trust the grouping. Click into a cluster from the dashboard to see what its members share, what each one *uniquely* knows, and pairwise cosine similarity. Members with no unique vocabulary stand out — those are your duplicates. **Topics** — tag-based groupings detected from skill names, descriptions, and a known-vocabulary scan. A skill can belong to multiple topics. Surfaced in the dashboard sidebar. ## Always-loaded vs on-use token cost Two different cost categories the dashboard separates: - **Always loaded** — the `description:` line at the top of every SKILL.md is loaded into Claude on **every session**, so the autorouter can decide which skill to invoke. You pay this whether you ever use the skill or not. Usually 30–100 tokens per skill. - **On use** — the body of the skill (everything below the frontmatter) loads **only when the Skill tool actually fires** for that skill. Usually 200–6,000 tokens. Both are graded separately because the rules are different. A bloated description hurts every session forever — that's the kind to trim aggressively. A bloated body only hurts the sessions where the skill actually fires, so depth there is fine if it pays off when triggered. ## Advanced: LLM-driven rewriting (opt-in) Default scan is deterministic. If you want machine-proposed specializations, pass `--rewrite` to run a 5-stage LLM pipeline: ```bash curator scan --rewrite # ~2–3 min per cluster curator review # walk through proposals as y/n flashcards ``` | Stage | Model | Purpose | |---|---|---| | 1. Detect clusters | none | Deterministic clustering | | 2. Fingerprint | Haiku 4.5 | Extract each skill's unique strengths | | 3. Specialize | Sonnet 4.6 | Propose renames + tightened prompts | | 4. Generate router | Sonnet 4.6 | Produce a router SKILL.md | | 5. Verify | Sonnet 4.6 (fresh instance) | Independent second opinion | This is a power-user mode, not the headline. The dashboard alone is the headline. ## How it compares `curator` is not the first tool in this space. Honest comparison in [`docs/COMPARISON.md`](./docs/COMPARISON.md). Short version: - [`claudetoolkit.com`](https://claudetoolkit.com/) — closed-source desktop GUI, freemium. Closest analog. Pick this if you want a polished commercial product covering every config type. We're the open-source alternative focused on skills + MCPs with clustering on top. - [`claude-config-audit`](https://github.com/paolodalprato/claude-config-audit), [`mcp-optimizer`](https://github.com/choam2426/mcp-optimizer), [`mcp-tidy`](https://github.com/nnnkkk7/mcp-tidy), `mcp-checkup`, [`claude-code-organizer`](https://github.com/mcpware/claude-code-organizer) — all open-source CLI auditors. None have a web dashboard or clustering. We borrow detection patterns from these (with credit). ## Roadmap **v0.1 (now)** - Local web dashboard, cluster overlap analysis, always-loaded vs on-use token grading - MCP duplicate + token-cost detection, orphan & drift flagging - Weekly cron snapshot - Optional `--rewrite` LLM pipeline **v0.2 (next)** - MCP server editing through the UI - Bulk select + delete - Slash commands and agents in the dashboard - Side-by-side skill compare - Effectiveness tracking from `~/.claude/projects/*.jsonl` (find skills that *should* have triggered but didn't) - Cross-tool: Cursor, Codex, OpenCode, Claude Desktop ## License & contributing MIT — see [LICENSE](./LICENSE). The CLI is and stays MIT. Contributions go through a CLA (see [CONTRIBUTING.md](./CONTRIBUTING.md)) — required to keep the project permissively licensed long-term. The CLA bot signs you in on your first PR.