An MCP server that turns Claude into a proactive skill curator.
It notices when you keep doing the same thing by hand — then offers the right skill.
No searching. No docs. Just say yes or no.

Install · How it works · AI protocol

   Claude Code
   ┌────────────────┐           ┌──────────────────────┐
   │ omakase-chef   │           │ claude-omakase MCP   │
   │ SKILL.md       │ calls ──▶ │ server (stdio)       │
   │ (the behavior) │           │ (registry + install) │
   └────────────────┘           └──────────────────────┘
                                           │
                                           │ fetched at build time
                                           ▼
   ┌─────────────────────────────────────────────────┐
   │ Daily federated catalog                         │
   │ ├── handpicked/        verified seeds + overlay │
   │ ├── skillsmp           public marketplace       │
   │ └── github-skills      GitHub SKILL.md search   │
   └─────────────────────────────────────────────────┘

How it works

You install the MCP server — one npx command added to your MCP host config.
Claude watches — the bundled omakase-chef skill instructs Claude to monitor your workflow in the background.
Claude notices — after you repeat the same kind of manual task three times, Claude surfaces the right skill from the catalog.
You approve — Claude installs it to ~/.claude/skills/ and it's active next session.

First time? Claude greets you with a starter pack of universally useful skills so you're not starting from zero.

Search that speaks your language. find_skill ranks the catalog by keyword plus LLM-generated synonyms in both English and Korean, then returns the top candidates and lets Claude rerank them by reading the descriptions — so a query like 웹사이트 크롤링 or "scrape a site" both land on the right web-scraping skill, even when its description is in the other language. Search only has to surface the answer somewhere in the top few; Claude (already in the loop) picks the best fit — no extra model at serving time.

Install

# 1. Install the omakase-chef skill
npx -y -p claude-omakase claude-omakase-install

# 2. Add the MCP server to your Claude Code config
claude mcp add omakase -- npx -y claude-omakase

Requirements: Node.js 20+, Claude Code (or another MCP host that supports stdio servers).

That's it. Restart your Claude session and the chef is active.

Manual MCP config (Claude Desktop or other hosts)

Add this to your MCP host config file and restart the host:

{
  "mcpServers": {
    "omakase": {
      "command": "npx",
      "args": ["-y", "claude-omakase"]
    }
  }
}

What Claude does for you

Once registered, these things just happen:

Situation	Claude does
You summarize three PRs in one session	Offers to install `pr-summarizer`
You mention "I always have to do X by hand"	Searches for a skill that covers X
You ask "what can I install?"	Runs `recommend_skills` against your profile
No skills installed yet	Shows the starter pack
No catalog match found	Drafts a new `SKILL.md` tailored to your task using MCP sampling

You can also be explicit: "I have a recurring task that doesn't have a skill yet" → Claude calls propose_new_skill, writes a draft SKILL.md to ~/.claude/skills/, and iterates with you until it's right.

Use cases

Concrete, end-to-end walkthroughs of what Omakase actually does.

1. First session — onboarding from zero

You:     (start a new Claude session, no skills installed)
Claude:  New here — want to set up a starter pack? I'll pop up a picker.
         → [native checkbox dialog appears]
            [x] Quick Review     — one-line, severity-tagged diff feedback
            [x] Understand Anything — deep explanations that lead with WHY
            [ ] Grill Me         — stress-test a plan by getting interviewed
            [ ] Write a Skill    — turn a workflow into a new skill
You:     (check two boxes, submit)
Claude:  Installed Quick Review and Understand Anything to ~/.claude/skills/. Active next session.

On clients with MCP elicitation (e.g. Claude Code), recommend_skills shows a real interactive checkbox dialog and installs exactly what you check — no typing, no parsing. The same tool drives this for both a brand-new user (the full pack) and a returning user with gaps (just the missing staples). On clients without elicitation it says so and returns the skills as a Markdown checklist for you to pick from.

2. Repeated manual work — proactive nudge

You:     (run `git diff`, review by hand; repeat across three PRs in one session)
Claude:  You've reviewed three diffs by hand this session. There's a Quick Review skill
         that gives one-line severity-tagged feedback — want me to install it?
You:     yes
Claude:  Done. Say "review this diff" next session to use it.

3. Intent in your prompt — suggestion without asking

You:     I need to automate browser testing across Chrome and Firefox
Claude:  That lines up with a browser-testing skill in the catalog (not installed yet) —
         want me to add it? (otherwise I'll just help directly)

4. Nothing matches — draft a new skill

You:     I keep reformatting our weird internal changelog format by hand
Claude:  No catalog skill covers that. Want me to draft one? I'll write a SKILL.md
         tailored to your changelog format and we'll iterate.

5. Manage what you've installed

npx claude-omakase tui

An interactive terminal app to list installed skills, health-check them
(missing SKILL.md? missing receipt? update available?), update them from the
catalog, and remove them — all with arrow keys and checkboxes. See
Manage your skills below.

Starter pack

When you have no skills installed, recommend_skills returns four universally useful skills:

Skill	What it does
Grill Me	Stress-tests your plans — Claude interviews you relentlessly, one hard question at a time
Understand Anything	Deep explanations of code, systems, or concepts — leads with WHY, not HOW
Write a Skill	Turns any recurring workflow into a reusable Claude skill
Quick Review	Fast, severity-tagged code review (`🔴 bug`, `🟡 perf`, `🔵 style`) — no praise, no fluff

Install any of them: "Install the Grill Me skill" → Claude calls install_skill("grill-me").

MCP tools

Tool	Description
`find_skill`	Search the catalog by task description (keyword + EN/KO synonym match, IDF-weighted); returns the top-K candidates for Claude to rerank and pick from
`list_installed_skills`	List installed skills and install receipts
`install_skill`	Download and install a skill to `~/.claude/skills/<id>/`
`recommend_skills`	Ranked suggestions based on your profile, context, and install state. With no context it drives starter-pack onboarding — full pack for a new user, missing staples for a returning one — via a real checkbox picker (MCP elicitation), or a Markdown checklist on clients without it
`offer_skill`	Offer one found skill with an interactive Install / Not now / Never recommend picker (MCP elicitation); "never" blocks it from future `find_skill`/`recommend_skills`
`uninstall_skill`	Remove `~/.claude/skills/<id>/` and its install receipt (idempotent)
`update_skill`	Re-download a skill's files from the catalog
`doctor_skills`	Health report per skill: SKILL.md present? receipt present? in catalog? version match?
`set_profile`	Save your role, languages, and tools so recommendations improve over time
`propose_new_skill`	Draft a new SKILL.md from scratch using MCP sampling

Manage your skills

A terminal UI for the day-to-day lifecycle, no Claude session required:

npx claude-omakase tui      # or: omakase tui

It shows a health dashboard for every installed skill and lets you act on them:

┌  🍱  Claude Omakase — skill manager
│
◇  3 installed · 2 healthy · 1 need attention ───────────╮
│       Skill                SKILL.md   Receipt  Catalog │
│   ──────────────────────────────────────────────────── │
│   ✅  quick-review         ✓          ✓        ✓       │
│   🔄  grill-me             ✓          ✓        update! │
│   ⚠️  understand-anything  ✗ missing  ✓        ✓       │
╰────────────────────────────────────────────────────────╯
│
◇  What would you like to do?
│  ● Re-run health check
│  ○ Update skill(s) from catalog
│  ○ Remove skill(s)
│  ○ Quit

Health check — flags skills with a missing SKILL.md, a missing install receipt, or a newer version available in the catalog.
Update — re-downloads selected skills' files from the catalog (multi-select).
Remove — deletes selected skills and their receipts, with a confirmation prompt.

The same operations are available to Claude as the doctor_skills, update_skill, and uninstall_skill MCP tools.

Proactive hooks (optional)

Three opt-in Claude Code hooks make discovery deterministic instead of relying on Claude noticing on its own. The installer copies them to ~/.claude/hooks/omakase/ (a stable path), but never registers them for you — you opt in by pasting the snippet it prints into your settings.json.

Hook	Event	What it does
`omakase-session-start.mjs`	`SessionStart`	At the start of a fresh session (`startup`/`clear`, not resume), tells Claude to run the chef's onboarding routine now — so new users get the starter pack without asking. Fires at most once per cooldown window.
`omakase-repetition.mjs`	`PostToolUse` (Bash + edits)	Tracks command workflows (single commands and multi-step chains via n-gram detection) and repeated edits to the same file (`edit:<name>`, counted once per session) in one cross-session file with timestamps; when a task recurs 3× within a rolling window, nudges Claude to find a matching skill. Primitive tools (`grep`, `cut`, `find`, …) and VCS/build plumbing (`git status`, `npm run`, `gh pr`, …) are filtered out, so it fires on real tasks, not on the commands you run to do them. Stays silent if no catalog is available.
`omakase-suggest.mjs`	`UserPromptSubmit`	Matches each prompt against the catalog; if a not-yet-installed skill clearly fits, suggests it once per session (with a cooldown).

{
  "hooks": {
    "SessionStart": [
      { "hooks": [
        { "type": "command", "command": "node ~/.claude/hooks/omakase/omakase-session-start.mjs" }
      ] }
    ],
    "PostToolUse": [
      { "matcher": "Bash|Edit|Write|MultiEdit|NotebookEdit", "hooks": [
        { "type": "command", "command": "node ~/.claude/hooks/omakase/omakase-repetition.mjs" }
      ] }
    ],
    "UserPromptSubmit": [
      { "hooks": [
        { "type": "command", "command": "node ~/.claude/hooks/omakase/omakase-suggest.mjs" }
      ] }
    ]
  }
}

Tunable via env vars: OMAKASE_SESSION_COOLDOWN_HOURS (default 24; set 0 to greet on every startup), OMAKASE_REPETITION_THRESHOLD (default 3), OMAKASE_REPETITION_WINDOW_DAYS (default 14), OMAKASE_SUGGEST_THRESHOLD (default 5), OMAKASE_SUGGEST_COOLDOWN (default 3 prompts). All are local-only — no network, no telemetry.

Catalog

The catalog is federated from multiple sources in CI. Adapters generate entries, and a daily CI job opens a PR if the committed catalog drifts. Runtime serving reads the committed/cached catalog; it does not scrape upstream sources on the user's machine.

Adapter	Source	Notes
`handpicked`	`handpicked/*.json` in this repo	Manually audited, `verified: true`
`skillsmp`	skillsmp.com	Public marketplace, expanded daily in CI via category + intent seeds
`github-skills`	GitHub code search	Public `SKILL.md` files, requires `GITHUB_TOKEN` during catalog builds

Only real Claude Code skills (entries with a resolvable SKILL.md) are kept — build:catalog --probe prunes anything that 404s. After federation, a build-time pass (enrich:catalog) asks an LLM for English + Korean search synonyms per skill and stores them on the entry, so search works in either language with no model at serving time.

Want to add a source? See src/adapters/README.md and run:

npm run scaffold:adapter -- <source-name>

Privacy

Everything Omakase writes stays on your machine:

Path	Contents
`~/.claude/skills/<id>/`	Installed skill files
`~/.config/claude-omakase/profile.json`	Your profile (role, languages, tools)
`~/.local/share/claude-omakase/installed/`	Install receipts
`~/.cache/claude-omakase/catalog.json`	Cached remote catalog, when `CLAUDE_OMAKASE_CATALOG_URL` is configured

No telemetry. No accounts. Outbound calls happen only for configured remote catalog refresh and skill file downloads (when you approve an install).

Trust & safety

Omakase installs third-party content into ~/.claude/skills/, so it treats the
catalog as a supply-chain surface. Two things are never automated: marking an
entry verified: true (human audit only) and writing a skill to disk (you approve
every install). Catalog entries pass three automated gates — schema, install
resolvability, HTTPS-only sources — before they're trusted or merged.

Full trust boundary, the verified definition, auto-merge guardrails, and the
rollback path: docs/TRUST.md.

Develop

git clone https://github.com/gyuch-an02/claude-omakase
cd claude-omakase
npm install
npm run build              # compile TypeScript
npm run build:catalog -- --probe   # fetch adapters → catalog.json, drop dead SKILL.md URLs
npm run enrich:catalog     # (optional) add EN/KO search synonyms via an LLM endpoint
npm run typecheck          # type check without emitting
npm test                   # run all tests

claude mcp add omakase -- node /path/to/claude-omakase/dist/server.js

Contributing

All contributions are welcome — from first-timers to seasoned TypeScript developers.

Good first issues: look for the good first issue label.

Quick ways to contribute:

New adapter: npm run scaffold:adapter -- <name> → implement fetch(): Promise<Entry[]> → register in src/adapters/index.ts → PR.
New handpicked entry: add a JSON file to handpicked/ following the Entry shape in src/types.ts.
Bug or feature: open an issue or PR — all contributions welcome.

See CONTRIBUTING.md for the full guide, including the PR checklist and local dev loop.

We follow the Contributor Covenant. Please read it before participating.

claude-omakase