> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cognizo.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP Server

> What the Cognizo MCP server is, and how it fits with the API.

The Cognizo **MCP server** lets an AI client call the Cognizo API on your
behalf — "how did Acme's visibility move this week", "which prompts are we
losing", "which domains cite our competitors but not us". It's a hosted,
remote [Model Context Protocol](https://modelcontextprotocol.io) server, so
there's nothing to install.

<CardGroup cols={2}>
  <Card title="Connect a client" icon="plug" href="/mcp/setup">
    Claude, Cursor, VS Code, Windsurf — one line each.
  </Card>

  <Card title="Tools" icon="wrench" href="/mcp/tools">
    All 42 tools: 38 API operations plus 4 workflow tools.
  </Card>
</CardGroup>

## At a glance

|               |                                                                 |
| ------------- | --------------------------------------------------------------- |
| **URL**       | `https://mcp.cognizo.ai/mcp`                                    |
| **Transport** | Streamable HTTP (over TLS)                                      |
| **Auth**      | Your Cognizo API key in the `X-API-KEY` header                  |
| **Tools**     | 42 — 38 generated from the public API, 4 hand-written workflows |
| **State**     | None. The server stores nothing between requests.               |

## How it works

```
Your AI client
   │  Streamable HTTP + TLS, X-API-KEY: cog_sk_…
   ▼
mcp.cognizo.ai/mcp
   │  tools generated from the public OpenAPI spec
   │  + workflow tools that chain several calls
   │  forwards YOUR key on every upstream request
   ▼
Cognizo public API  →  { data, meta }
```

The tool surface is generated from the same
[public OpenAPI spec](/api-reference/introduction) that powers the API
reference, so the tools and the endpoints never drift apart: one new public
endpoint means one new tool, named after the same operation id.

On top of that sit four **workflow tools** — hand-written shortcuts that chain
several API calls into the answer a question actually needs, so the model
doesn't have to orchestrate three round-trips to say how a brand is doing.

## What it can reach

Exactly what your API key can reach, and nothing more:

* The [public API surface](/api-reference/introduction) — brands, topics,
  prompts, citations, competitors, ads.
* Only brands your organization owns.

Anything outside that surface is rejected upstream with `403`, the same as a
direct API call. The MCP server has no privileges of its own — it is a proxy,
not an account.

<Note>
  The server holds **no** credentials. Your key travels with each request from
  your client, is forwarded upstream, and is never persisted. See
  [Security & troubleshooting](/mcp/troubleshooting).
</Note>

## Read and write

The generated tools cover writes as well as reads — creating brands, bulk-adding
prompts, tracking competitors, soft-deleting topics. An assistant with your key
can change your workspace, so treat MCP access like handing over dashboard
access.

If you want a read-only setup, issue a key you can revoke independently and
keep an eye on what your client is allowed to call — most clients let you
approve tools individually.