MCP server

Let Claude, Cursor and other AI assistants read your Notable data through the Model Context Protocol.

The Notable MCP server lets an AI assistant read your brands' visibility for you: "how did we do on Perplexity this month?", "which sites cite our competitors but not us?", "what should I work on next?". It returns the same data as the REST API and uses the same API key.

Server

https://www.heynotable.com/api/mcp

Transport: Streamable HTTP. Authentication: the header Authorization: Bearer ntb_… on every request.

Connect a client

Claude Code

claude mcp add --transport http notable https://www.heynotable.com/api/mcp --header "Authorization: Bearer ntb_…"

Claude Desktop, Cursor and others

Any client that takes a server URL and custom headers works. Most read a JSON configuration like this:

{
  "mcpServers": {
    "notable": {
      "type": "http",
      "url": "https://www.heynotable.com/api/mcp",
      "headers": { "Authorization": "Bearer ntb_…" }
    }
  }
}

Clients that can only sign in with OAuth, such as claude.ai custom connectors, are not supported yet.

Tools

Tool	Does	Arguments
`list_brands`	Every brand you can see. Start here to get brand ids	none
`get_brand`	Market, competitors and intents, with their ids	`brandId`
`get_metrics`	The four metrics for the brand and each competitor, latest 28 days	`brandId`, optional `intent`, `engine`
`get_trend`	One metric over time	`brandId`, `metric`, `days` (7–365)
`get_sources`	Cited domains, and sites cited for competitors but not you	`brandId`, optional `intent`, `engine`
`list_actions`	What to work on next, highest priority first	`brandId`, optional `status`
`set_action_status`	Mark an action `done`, `dismissed` or `todo`	`brandId`, `actionId`, `status`

Every tool except set_action_status is read-only, and is marked that way so clients can run it without asking. set_action_status needs a role that can edit the brand.

Results are JSON, with every rate carrying its 95% interval and sample size, as described in reading the numbers. A brand the key cannot see comes back as a tool error the assistant can read, not as a failed request.

Limits

The same as the API: 120 requests a minute per key. An assistant answering one question usually makes two to five calls.

Example prompts

"List my Notable brands, then compare our pick rate with each competitor's on ChatGPT."
"Which third-party sites get cited when competitors are recommended and we aren't?"
"Show my open Notable actions for Acme and mark the robots.txt one done."