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

# MCP Connector

> Connect any AI assistant to Comp AI with one URL and a sign-in — no install, no API key. The hosted, recommended way to use the Comp AI MCP.

The Comp AI **MCP Connector** is our hosted MCP server at `mcp.trycomp.ai`. You connect an AI assistant — Claude, Cursor, ChatGPT, VS Code — by pasting **one URL** and signing in. There is nothing to install and no API key to manage: access is granted through OAuth, scoped to your own Comp AI permissions, and revocable at any time.

> This is the recommended way to connect. If you'd rather run the server yourself from an API key (for a script, a CI job, or a client that can't use a remote connector), see [MCP Server](/docs/mcp-server) — the self-managed option.

## Why the connector

<CardGroup cols={2}>
  <Card title="Nothing to install" icon="bolt">
    No npm package, no config file. Paste a URL; the assistant handles the rest.
  </Card>

  <Card title="Sign in, don't copy a key" icon="key">
    OAuth in your browser. No API key to generate, paste, or rotate.
  </Card>

  <Card title="Your own permissions" icon="shield-check">
    The connection can only do what your Comp AI role can do — per user, not a shared key.
  </Card>

  <Card title="Revoke in one click" icon="ban">
    Turn off a connection from Settings whenever you want; it stops working immediately.
  </Card>
</CardGroup>

## The one thing you need

```
https://mcp.trycomp.ai/mcp
```

Paste that into any client's custom / remote MCP connector setup. Everything else — discovery, sign-in, approval — happens automatically.

<Note>
  The assistant connects to `mcp.trycomp.ai` from the vendor's cloud (this is true even for desktop apps), so nothing runs on your machine and the server must simply be reachable over the internet — which it is.
</Note>

## Connect your assistant

<Tabs>
  <Tab title="Claude (Desktop & web)">
    Custom connectors work on Free, Pro, Max, Team, and Enterprise plans.

    <Steps>
      <Step title="Open connector settings">
        Go to **Customize → Connectors**.
      </Step>

      <Step title="Add a custom connector">
        Click **“+” → Add custom connector**, then enter:

        * **Name:** `Comp AI`
        * **URL:** `https://mcp.trycomp.ai/mcp`

        Click **Add**. Leave *Advanced settings* empty — no client ID or secret is needed.
      </Step>

      <Step title="Sign in">
        Follow the sign-in and approval screen (see [What approval looks like](#what-approval-looks-like)).
      </Step>

      <Step title="Enable it in a chat">
        In a conversation, click the **“+”** button, choose **Connectors**, and enable **Comp AI**.
      </Step>
    </Steps>

    On **Team / Enterprise**, an organization Owner adds it once under **Organization settings → Connectors → Add → Custom → Web**; members then **Connect** it from **Customize → Connectors**.
  </Tab>

  <Tab title="Claude Code (CLI)">
    ```bash theme={null}
    claude mcp add --transport http comp-ai https://mcp.trycomp.ai/mcp
    ```

    Then start (or restart) Claude Code and run:

    ```
    /mcp
    ```

    Select **comp-ai** and complete the browser sign-in. That's it — no API key.
  </Tab>

  <Tab title="Other clients">
    Any assistant that supports a **remote / custom MCP connector over HTTP** works the same way — there is nothing Comp AI-specific to install. Add a custom connector and paste:

    ```
    https://mcp.trycomp.ai/mcp
    ```

    For the exact menu path, follow your client's own connector guide:

    * **Cursor** — Settings → MCP / Connectors → add a remote server
    * **ChatGPT** — Settings → Connectors → add a custom connector (availability varies by plan)
    * **VS Code** — the MCP: Add Server command, choosing an HTTP server

    If your client can't use a remote connector, use the [self-managed MCP Server](/docs/mcp-server) with an API key instead.
  </Tab>
</Tabs>

## What approval looks like

The first time you connect, you'll see an approval screen. It's built to make one thing obvious: **where your access is going.**

* **The destination host is shown large** — for the real connector it reads `mcp.trycomp.ai`. This is the field that matters; an app's *name* is chosen by whoever registered it, so the host is what to check.
* **Choose your organization.** If you belong to more than one, pick which one the connection acts on. (You can change it later in Settings.)
* **Tick to acknowledge**, then **Allow access**. The grant matches your own permissions — there is no read-only mode.

<Note>
  You may see “Comp AI has not reviewed this app.” That's expected for connectors we haven't formally vetted, and it isn't a failure — only continue if you started the connection yourself.
</Note>

## Manage or disconnect

Open **Settings → your profile → Connected apps**. Each connected assistant is listed with the host it sends tokens to and a **Revoke** button. Revoking deletes the grant and its tokens, so that assistant stops working immediately — reconnect from the assistant to use it again.

## Connector vs. self-managed

<CardGroup cols={2}>
  <Card title="MCP Connector (this page)" icon="cloud">
    Hosted by Comp AI. One URL, OAuth sign-in, per-user permissions, revoke in Settings. Best for people using an AI assistant day to day.
  </Card>

  <Card title="MCP Server (self-managed)" icon="terminal" href="/docs/mcp-server">
    You run it from a Comp AI API key via `npx`. Best for scripts, CI, or clients that can't use a remote connector.
  </Card>
</CardGroup>

## Resources

* [Model Context Protocol](https://modelcontextprotocol.io/) — the open standard this is built on
* [MCP Server](/docs/mcp-server) — the self-managed, API-key option