recomention

@recomention/mcp

An MCP server for recomention: ask Claude Code, Cursor or Codex *"is my product recommended by ChatGPT for meeting scheduling?"* and have it start a report, wait for it, and read the result — every engine's "named in n of m answers", who is recommended instead, the sources the engines read, and the fix list.

It is a thin stdio client for the public JSON API v1 at https://api.recomention.com/v1. It holds no state. It runs locally, as a process your MCP client starts. For ChatGPT, Claude.ai and other clients that connect to a remote server, use the hosted connector at https://mcp.recomention.com/mcp instead: it signs in with your recomention account (OAuth), no API key needed. Setup for each client: https://docs.recomention.com/mcp/.

Requires Node.js 20 or newer (npx comes with it).

Environment

Install

Claude Code

claude mcp add recomention \

--env RECOMENTION_API_KEY=YOUR_API_KEY \ --env RECOMENTION_BASE_URL=https://api.recomention.com \ -- npx -y @recomention/mcp

Replace YOUR_API_KEY with your key. Add --scope user to use it in every project. Check it with claude mcp list, or /mcp inside a session.

Cursor (~/.cursor/mcp.json, or .cursor/mcp.json in a project) and other JSON-configured clients

{

"mcpServers": { "recomention": { "command": "npx", "args": ["-y", "@recomention/mcp"], "env": { "RECOMENTION_API_KEY": " ", "RECOMENTION_BASE_URL": "https://api.recomention.com" } } } }

Keep a project-level .cursor/mcp.json that holds a key out of version control.

Codex

codex mcp add recomention \

--env RECOMENTION_API_KEY=YOUR_API_KEY \ --env RECOMENTION_BASE_URL=https://api.recomention.com \ -- npx -y @recomention/mcp

or in ~/.codex/config.toml:

[mcp_servers.recomention]

command = "npx" args = ["-y", "@recomention/mcp"] env = { RECOMENTION_API_KEY = " ", RECOMENTION_BASE_URL = "https://api.recomention.com" }

Tools

Reading results honestly: an engine in enginesSkipped did not run (for example *"provider rejected the key"*) — it is never an engine that left the product out. A report with simulated: true is test data.

Setup prompt

The account page's Copy setup prompt button copies this (with the site's own URL). Paste it into your agent:

Set up the recomention MCP server so you can check whether AI engines (ChatGPT, Claude, Perplexity, Gemini, Grok) recommend a product.

1. Add this MCP server to your MCP configuration (Claude Code: claude mcp add; Cursor: ~/.cursor/mcp.json; Codex: ~/.codex/config.toml): command: npx -y @recomention/mcp env: RECOMENTION_API_KEY= RECOMENTION_BASE_URL=https://api.recomention.com Claude Code one-liner: claude mcp add recomention --env RECOMENTION_API_KEY= --env RECOMENTION_BASE_URL=https://api.recomention.com -- npx -y @recomention/mcp
2. Restart or reload MCP servers, then call recomention_rankings_list to confirm it works.

How to use it:

• To check a product, first call recomention_suggest_prompts with its domain, category and a one-line description, show me the questions, and let me edit them.
• Starting a report spends money: an agency credit, or calls on my own provider keys (that needs my Agent plan). Always ask me before calling recomention_report_start, and say which funding it will use.
• On my own keys, any subset works, but one of OpenAI, Anthropic or Gemini is required. If the start is refused with a limit (plan, projects, reportsPerMonth, perDay, concurrent), tell me which limit and do not retry around it.
• Reports take several minutes. Poll recomention_report_status about once a minute; do not start a second report for the same product while one is running.
• When it is COMPLETE, call recomention_report_get and summarise: the overall score, each engine's "named in n of m answers", who is recommended instead, and the top three fixes. Quote counts as "n of m", never as yes/no.
• An engine listed under enginesSkipped did not run (for example "provider rejected the key"). Say so; never describe it as an engine that did not mention the product.
• A report with simulated: true is a test run with made-up answers. Say that plainly and do not draw conclusions from it.

Docs: https://www.recomention.com/developers

The canonical text lives in packages/core/src/config/developers.ts (setupPrompt); keep this copy in step with it.

Skills

skills/ holds agent skills in the slash-command style: copy a folder into .claude/skills/ (Claude Code) or your agent's equivalent.

• ai-visibility-check — check one product end to end.
• fix-list-review — turn a finished report into a prioritised, honest plan of work.
• competitor-watch — compare who the engines name in a category, from public rankings and your reports.
• distribution-kit — read a finished report's kit and work through its channel plan, one channel at a time.

Develop

npm install

npm test # in-memory client against the server, fake API; metadata/version checks npm run smoke # spawns the stdio server against RECOMENTION_BASE_URL npm run build # clean dist/ (bin: recomention-mcp) npm run smoke:pack # packs the tarball (with pnpm), checks its contents, runs it against a fake local API

In the monorepo, run them as pnpm --filter @recomention/mcp .

Run smoke --start only against a local server with RECOMENTION_MOCK=1: against a live deployment it spends a credit.

prepack empties and rebuilds dist/, so a pack or publish always ships the current sources and no test files. version lives in package.json; the server reads it at startup, and server.json (the MCP Registry entry, io.github.recomention/mcp) must carry the same number — a test fails otherwise.

This repository is published from the recomention monorepo, where development happens. Bugs and questions: issues or support@recomention.com.

Licence

MIT.