Back to search
skill-router
MarketplaceWorks with Claude CodeGitHub
skill-router: picks the right skills, plugins and MCP servers for every Claude Code message.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/fripouilledu92--skill-router)HTML
<a href="/listings/fripouilledu92--skill-router"><img src="/api/listings/fripouilledu92--skill-router/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
claude plugin marketplace add Fripouilledu92/skill-router claude plugin install skill-router@skill-router
Open a new Claude Code session. Your tools are classified once in the background, then every message is routed locally in about 0.1 s, with no model call.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 14 lines
- {
- "name": "skill-router",
- "description": "skill-router: picks the right skills, plugins and MCP servers for every Claude Code message.",
- "owner": {
- "name": "Fripouilledu92"
- },
- "plugins": [
- {
- "name": "skill-router",
- "source": "./",
- "description": "Picks the skills, plugins and connected MCP servers that fit each message, in the right order, and explains why."
- }
- ]
- }
README.md · README · 108 lines
- <p align="center"><img src=".claude-plugin/icon.png" width="96" alt="skill-router icon"></p>
- # skill-router
- **Your Claude Code agent has 100+ skills, plugins and MCP servers. It uses three.**
- skill-router picks the right ones for every message, hands them to Claude in the order a pro would use them,
- and Claude ends each answer with one line saying what each tool brought.
- <p align="center"><img src="assets/demo.png" alt="skill-router choosing brainstorming, a design skill, unslop, break-ui and deploy for a landing page request"></p>
- <p align="center"><sub>Illustration: the exact list depends on the tools you have installed.</sub></p>
- ## Install
- ```bash
- claude plugin marketplace add Fripouilledu92/skill-router
- claude plugin install skill-router@skill-router
- ```
- Open a new Claude Code session. Your tools are classified once in the background, then every message is routed
- locally in about 0.1 s, with no model call.
- ## How it works
- | When | What happens | Cost |
- |---|---|---|
- | Install / new session | Every skill, agent, command, plugin and **connected** MCP server is listed. New or changed ones are classified once, in the background, by Claude: category, tags, specialty, phase, neighbors. | One `claude -p` call per batch of new tools, once |
- | Every message | A local script (no model call) scores your message against the index and gives Claude a shortlist, the usual chain for that kind of task, and the nuances between similar tools. | 0 model calls, ~250-400 tokens of context, ~0.1 s |
- | Claude answers | Claude keeps the tools that really help, uses them in order (frame → design → build → verify → ship) and ends with a `Tools:` line. | — |
- | After the answer | It notes which proposed tools were actually used; useful ones climb, ignored ones sink a little. | 0 |
- **Neighbors, not duplicates.** Two design skills are not the same skill. Each tool carries a specialty and a
- "when to pick me over my neighbor" nuance, so Claude chooses `unslop` to clean an AI-looking screen and
- `impeccable` to design one from scratch.
- **Only what is connected.** MCP servers that need auth or failed to start are ignored (`claude mcp list`).
- **Quiet on small talk.** "ok", "thanks", "continue" and slash commands trigger nothing.
- ## Commands
- | Command | What it does |
- |---|---|
- | `/why` | Why these tools were proposed for the last message, which words matched, which were really used |
- | `/skills-map` | Your whole toolbox, by category, with security notes |
- | `/skills-off <name>` · `/skills-off on <name>` | Never propose a tool again / bring it back |
- | `/skills-sync` | Classify new tools now, including the ones only the agent can see (desktop app connectors) |
- ## Requirements
- Claude Code, Python 3.9+, the `claude` CLI on your PATH. The first classification of ~150 tools takes a few
- minutes in the background; until then tools are matched on their raw descriptions.
- ## Settings
- `~/.claude/skill-router/prefs.json`:
- ```json
- { "max": 5, "modele": "sonnet", "langue": "english", "classement_auto": true, "off": [] }
- ```
- ## Fixing a classification
- Anything the automatic classification got wrong goes in `~/.claude/skill-router/corrections.json`. Corrections
- always win over the classifier and survive re-classification:
- ```json
- {
- "categories": { "Old name": "New name" },
- "items": { "skill:unslop": { "phase": "verifier", "specialite": "…" } },
- "recettes": { "Design": [ { "quand": "…", "mots": ["site"], "chaine": ["skill:a", "skill:b|skill:c", "skill:d"] } ] }
- }
- ```
- `"skill:b|skill:c"` in a chain means "one of these, pick by the nuances".
- ## Security
- - Classification only **reads** tool files. It never runs a skill's scripts.
- - Each tool gets an `alerte` field when it sends data out, publishes, spends money or touches your config;
- `/skills-map` shows them.
- - The router only ever *suggests*. Your permission settings still apply to everything Claude does.
- ## Limits (v0.1)
- - Claude Code only. Codex, Cursor and Gemini CLI adapters are planned.
- - claude.ai on the web has no hooks, so it cannot run there automatically.
- - Matching is lexical (with FR/EN synonyms) for speed and zero cost: it builds a generous shortlist and lets
- Claude make the final call.
- - Learning ignores claude.ai connectors (their tool names are opaque ids).
- ## Tests
- The cases are examples from the author's own setup: they check the router against *your* index
- (`~/.claude/skill-router/index.json`), so adapt `tests/cas.json` to the tools you have.
- ```bash
- python3 tests/evaluer.py # labelled cases used while tuning
- python3 tests/evaluer.py tests/cas-inedits.json # cases written after tuning
- ```
- ---
- **En français** : à chaque message, skill-router choisit les skills, plugins et serveurs MCP connectés qui servent
- la demande, les donne à Claude dans l'ordre, et Claude termine sa réponse par une ligne « Outils : … ». Le tri se
- fait une fois à l'installation et à chaque ajout ; l'analyse de chaque message est locale et ne coûte aucun appel
- au modèle.
- MIT License.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.