Allergen Detector

Allergen Detector MCP

The Allergen Detector matching engine as an MCP server, so AI assistants and shopping agents (ChatGPT, Claude, and others) can check a product against a household's allergies before recommending or buying it.

It uses the same matcher.js and allergen map as the Chrome extension (51 allergens, 3,446 names in 24 languages). The map refreshes daily from GitHub, so map updates reach agents with no redeploy.

A short word that is an allergen in one language and an everyday word in another ("ou", "malta") only counts on text in its own language. The server infers the language from the "Ingredients" label word and, on text of 12 words or more, a detector; agents can also pass language. Short text with no label is not filtered. Words set aside are returned in other_language_words_not_counted, and languages_assumed shows the inference.

Tools

The same checks are also a plain web API for agents that don't use MCP: POST /api/check, POST /api/lookup and GET /api/allergens, described at /openapi.json.

API keys are optional. Set API_KEYS="muse= ,acme= " on the host to give each partner its own key, sent as X-API-Key or Authorization: Bearer. Requests without a key still work; a wrong key gets a 401. A valid key names the partner in /stats.

All tools are read-only and stateless. Nothing about the person (allergens, ingredients) is stored or logged. The only log is an anonymous count of tool calls per day, for measuring traction.

Run it

npm install

npm test # end-to-end test over MCP npm start # remote server: http://localhost:3000/mcp npm run stdio # local server for Claude Desktop and similar npm run sync # copy the latest matcher.js and map from the extension

To use it locally in Claude Desktop, add this to the mcpServers section of its config:

"allergen-detector": { "command": "node", "args": ["/path/to/allergen-mcp/server.js", "--stdio"] }

Launch steps

1. Host it at a public HTTPS URL. Render or Fly.io can run npm start from a GitHub repo for $0–7/month. Later, point mcp.safi.ai at it.
2. Privacy policy and terms page. Every directory requires one. Until safi.ai is live, host it on GitHub Pages. Its main point: no health data is stored.
3. List it in the official MCP Registry (registry.modelcontextprotocol.io) with the mcp-publisher CLI, under io.github.daniellemeng1022/allergen-detector. Many clients and directories pull from it.
4. Submit to the community directories: Smithery, Glama, PulseMCP and mcp.so.
5. Claude: anyone can add it today as a custom connector by pasting the URL. To appear in Anthropic's connectors directory, submit it through their directory form.
6. ChatGPT: test it in developer mode, then submit it as an app through OpenAI's Apps SDK submission process. This needs a verified developer account.
7. Meta Muse (do this first): submit at muse.ai/platform. Choose "Existing MCP," enter the hosted URL, choose "API keys" and give Meta a key from API_KEYS (e.g. muse=…) so its calls show up as muse in /stats and payments to "no," then describe the service. Meta tests the live endpoint, so step 1 has to be done first. Muse already shops at Sephora, Walmart, Best Buy and the Shopify catalog, which covers both of our categories (cosmetics and food).
8. Instinct and similar assistants: Instinct doesn't document custom MCP support. It shops through Shopify and Shop Pay. Reach out for a partnership instead.
9. Shopping agents: contact teams building agentic-commerce products and offer the endpoint free during a pilot.
10. Measure: set STATS_KEY on the host, then open https:// /stats?key= to see calls by day, agent and tool. A connect count is an agent connecting; tool counts are real checks. /stats resets when the server restarts, but every count is also a [usage] JSON line in the host's logs. Track weekly checks and the number of distinct agents.

Before scaling

• Add rate limiting and an API key for heavy business users. Keep a free tier.
• Keep the disclaimer in every response. Don't present results as medical advice.