MCP server: connect Rankion to Claude and AI agents

Rankion runs its own MCP server (Model Context Protocol). Every tool call goes through the matching endpoint of REST API v1, so team access, validation and credits work exactly as they do for a direct API call.

Create the token in Rankion under Settings, section API-Tokens (https://rankion.ai/settings). See also Authentication.

Setup in Claude Code

claude mcp add --transport http rankion https://rankion.ai/mcp \
  --header "Authorization: Bearer <TOKEN>"

Replace <TOKEN> with your token. claude mcp list then shows the server as connected. With a wrong or revoked token the server answers 401 and the connection fails.

Setup for Managed Agents

Managed Agents is the agent API of the Claude Developer Platform. Setup has two parts: the agent names the server, and the token sits separately in a vault.

1. Add the server to the agent. Every entry in mcp_servers needs a matching mcp_toolset:

{
  "mcp_servers": [
    { "type": "url", "name": "rankion", "url": "https://rankion.ai/mcp" }
  ],
  "tools": [
    { "type": "mcp_toolset", "mcp_server_name": "rankion" }
  ]
}

2. Store the token as a vault credential of type static_bearer:

{
  "display_name": "Rankion",
  "auth": {
    "type": "static_bearer",
    "mcp_server_url": "https://rankion.ai/mcp",
    "token": "<TOKEN>"
  }
}

3. Pass the vault when you start the session: "vault_ids": ["<VAULT_ID>"].

Credentials are matched by URL, so mcp_server_url must point to the same server as url. The full calls are in Anthropic's documentation on MCP in Managed Agents and on vaults.

claude.ai and Claude Desktop

The one-click connector in claude.ai (web) and Claude Desktop requires OAuth sign-in. The Rankion server cannot do that today: it accepts bearer tokens only, and laravel/passport, the package for OAuth sign-in, is not installed. This is planned as phase 2, with no date set.

So there is no ready-made connection for these two surfaces yet.

The catalog updates itself

The tool list is rebuilt from the current v1 routes on every connection. A new route becomes a tool without any further setup.

The version field in serverInfo has the form 1.<tool-count>.<hash>, for example 1.747.dd625a1e. It changes whenever the API changes. A client must not cache the catalog across a version change. If version changes or an expected tool is missing, it re-reads the tools or reconnects.

The tool list is large

As of 2026-11-04 there are 747 tools. Of 754 endpoints, 744 have a tool, and three endpoints have two tools with separate descriptions. The current number is always in version.

This has two consequences:

• A client with a tool limit may not see all tools. Where possible, enable only the tools you need. In Managed Agents you do this in the mcp_toolset: "default_config": { "enabled": false }, plus the tools you want in configs with "enabled": true.
• tools/list returns 15 tools per page. A custom client has to follow nextCursor, otherwise it only knows the first 15.

All endpoints are listed at https://rankion.ai/api-reference.

These 10 routes have no tool

Files and inbound webhooks do not fit into a tool call. These routes are excluded on purpose (as of 2026-11-04) and stay available through the REST API. Other export routes do have their own tool.

Skill, REST API and OpenAPI

The MCP server adds to the other routes, and one token works for all of them:

• Skill for Claude Code: a Markdown file that Claude Code reads directly.
• REST API: all endpoints under https://rankion.ai/api/v1/..., described in the API reference.
• OpenAPI spec: machine-readable at https://rankion.ai/api/v1/openapi.yaml, for example for n8n, Zapier or custom clients.

Install the skill:

mkdir -p ~/.claude/skills/rankion-api
curl -o ~/.claude/skills/rankion-api/SKILL.md https://rankion.ai/claude-skill
export RANKION_API_TOKEN="<TOKEN>"

Test the REST API:

curl -H "Authorization: Bearer $RANKION_API_TOKEN" \
  -H "Accept: application/json" \
  https://rankion.ai/api/v1/team

Things to watch

• Never commit the token. Keep it in a .env file or a secret manager.
• Keep an eye on credits. A tool call costs as much as the endpoint behind it.
• Poll long-running jobs. Some endpoints answer with HTTP 202 and a job ID. Check the status afterwards with the matching tool or GET endpoint.

Related wiki articles

• API reference
• Authentication
• Master Agent / Agentic Chat: multi-step tasks via POST /v1/agentic/chat