Connect Your API Docs to Claude Desktop via MCP

Claude Desktop is most developers' daily AI driver. It's fast, the interface is clean, and its tool-use support — via the Model Context Protocol (MCP) — means you can wire it up to any system that speaks the protocol. Your API docs should be one of them.

When you connect Outworx Docs' per-project MCP server to Claude Desktop, Claude gains twelve new capabilities: listing and searching endpoints, fetching full schemas, generating TypeScript types, executing live API calls with your credentials, and running grounded Q&A over your spec. You stop pasting curl commands into chat. You stop explaining your API at the start of every conversation. Claude just knows.

This guide walks through the whole setup — five minutes, one config file.

Prerequisites

• A project hosted on Outworx Docs (free signup, Pro or Business plan for MCP access)
• Claude Desktop installed on macOS, Windows, or Linux
• Your project's MCP token (grab it from your dashboard in a moment)

Step 1: Get your MCP endpoint + token

Open your Outworx Docs dashboard, click into any project, and go to the MCP tab. You'll see two values:

• Endpoint URL — looks like https://docs.outworx.io/api/mcp/your-slug
• Access token — starts with mcp_, about 50 characters

Click the eye icon to reveal the token, then copy both values. If your project is on the Free plan you'll see an upgrade gate instead — MCP is a Pro / Business feature.

Step 2: Open Claude Desktop's config

The config file lives at:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

If it doesn't exist yet, create it. If it does and you already have other MCP servers wired up, you'll add a new entry alongside them.

Step 3: Paste the config

Replace <YOUR_SLUG> and <YOUR_TOKEN> with your values:

{
  "mcpServers": {
    "outworx-my-api": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-fetch",
        "https://docs.outworx.io/api/mcp/<YOUR_SLUG>",
        "--header",
        "Authorization: Bearer <YOUR_TOKEN>"
      ]
    }
  }
}

Claude Desktop uses a local stdio bridge (@modelcontextprotocol/server-fetch) to talk HTTP to remote MCP servers. The -y flag lets npx auto-install the bridge on first run. If you've already got it installed globally you can replace npx -y @modelcontextprotocol/server-fetch with mcp-server-fetch.

The MCP tab in your dashboard has this snippet pre-filled with your actual token — paste directly from there to skip the manual edit.

Step 4: Restart Claude Desktop

Fully quit the app (⌘Q on macOS — not just close the window) and reopen. Claude Desktop only reads the MCP config on startup.

When it reloads, you should see a small plug icon near the input box. Click it — the dropdown lists "outworx-my-api" with a green dot. That's your server, connected.

Step 5: Use it

Start a new chat and try something like:

Using the outworx-my-api tools, list every endpoint tagged "posts" and generate TypeScript types for the first three.

Claude will:

1. Call list_endpoints with tag="posts" to get the endpoint list
2. Call generate_types with scope="endpoint" for each one
3. Produce a single self-contained TypeScript module

No copy-paste from the docs site. No re-explaining what the API does. The schemas, descriptions, and examples are all just there.

What else Claude can do with this connection

Twelve tools are exposed through the MCP server. The highlights:

• ask_docs — natural-language Q&A grounded only in your spec. Claude uses this before making up answers about your API.
• execute_endpoint — hit your API live with caller-provided auth. Claude samples the response, and Outworx infers the schema automatically so future type generation reflects reality even when your OpenAPI is incomplete.
• search_endpoints — semantic search across every endpoint and schema. Useful when Claude doesn't know which endpoint to pick.
• recheck_types — validate that an endpoint's response still matches the stored inferred schema. Great for pre-ship drift checks.

You don't have to memorize tool names. Claude will pick the right ones when you ask normal questions ("what's the simplest way to create a new user?") and the system prompt from the MCP server nudges it toward grounded, cite-aware answers.

Permissions and security

The access token in the config grants all MCP tools by default. If you want to scope it — say, disable execute_endpoint because you don't want Claude making real API calls — head to the Tools sub-tab in the MCP dashboard and toggle individual capabilities.

Tokens are opaque, revocable, and tied to one project. If you ever think one leaked, click Rotate on the dashboard: the old token stops working instantly and you'll need to update the config on any machine using it.

Troubleshooting

• Plug icon doesn't appear: usually a JSON syntax error in the config. Run it through jq . to validate.
• Tools show "Error: 401": token is wrong, expired, or the project's MCP is disabled. Re-copy from the dashboard.
• Tools show "Error: 402": owner's plan was downgraded to Free. Upgrade to Pro to re-enable.
• Timeouts on execute_endpoint: the API you're targeting is slow or unreachable. Rate-limited to 15 calls/minute per project.

Once it's wired up, it tends to stay wired up. The workflow unlocks a lot — give it an afternoon and you'll wonder how you shipped integrations without it.