Annolux
EN 中文

MCP · Cursor

Cursor web search MCP setup

Add the search_web tool to Cursor's agent: one JSON file, a curated English and Chinese index, and fetched_at on every result.

Project or global

  • .cursor/mcp.json — this repository only; commit it to share the server definition (not the key).
  • ~/.cursor/mcp.json — every project on this machine.

The JSON

Install the annolux-mcp binary first (see the MCP overview below), then add this block. Replace the key placeholder with a real key.

{
  "mcpServers": {
    "annolux": {
      "command": "annolux-mcp",
      "env": {
        "ANNOLUX_API_URL": "https://api.annolux.com",
        "ANNOLUX_API_KEY": "ann_live_REPLACE_ME"
      }
    }
  }
}

Enable and approve

Open Cursor Settings → MCP: the annolux server should be listed with search_web as its tool; toggle it on if needed. In Agent mode the model calls the tool on its own; by default each call asks for your approval before it runs, so you always see the query being sent.

A concrete ask that works well

Try: "What is the current way to define props with TypeScript in Vue? Search with domains limited to vuejs.org." The agent sends one search_web call, gets ranked results from the official docs with fetched_at on each, opens the right page, and answers with a citation you can check. One successful call, one credit.

Point it at documentation you trust

The domains allow-list is the most useful field in an editor workflow: ask the agent to search with domains restricted to the library's own documentation site, and results stay inside the family you trust. Keep fetched_at next to any citation the agent writes into your code or notes.

Troubleshooting

  • Binary not found: use the absolute path from `which annolux-mcp` in the command field.
  • Edited the JSON but nothing changed: refresh the server list in Settings → MCP or restart Cursor.
  • Empty results are a normal outcome of a curated index — not an error to retry.

Boundaries and credits

search_web queries a curated English and Chinese index — not realtime, not the whole web. Only a successful 2xx search costs one credit; errors and timeouts cost zero. New accounts start with 1,000 free credits.