Model Context Protocol (MCP)¶

Rune supports the Model Context Protocol in both directions:

MCP Client — agents can connect to external MCP servers and use their tools
MCP Server — Rune exposes all deployed agents as MCP tools at POST /mcp

MCP Server — using Rune from Claude Desktop / Cursor¶

Any MCP-compatible client can connect to the Rune gateway and call agents as tools.

Endpoint: POST http://localhost:8080/mcp

Supported methods:

Method	Description
`initialize`	Handshake — returns `protocolVersion`, `capabilities`, `serverInfo`
`tools/list`	Returns all active deployed agents as MCP tools
`tools/call`	Invokes an agent by name with `{ input, session_id? }`
`ping`	Health check

Configure in Claude Desktop¶

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "rune": {
      "url": "http://localhost:8080/mcp",
      "type": "http"
    }
  }
}

After restarting Claude Desktop, all your deployed Rune agents appear as tools.

Configure in Cursor / VS Code¶

Add to your MCP settings:

{
  "mcp": {
    "servers": {
      "rune": {
        "url": "http://localhost:8080/mcp",
        "type": "http"
      }
    }
  }
}

MCP Client — agents consuming external MCP servers¶

Agents can connect to external MCP servers and use their tools. Declare them in the Runefile under mcp_servers::

name: my-agent
version: 0.1.0
instructions: |
  Use the filesystem and GitHub tools to help the user.
default_model: default
toolset:
  - rune@memory-store        # built-in
  - filesystem/read_file     # from MCP server "filesystem"
  - filesystem/write_file
  - github/create_issue      # from MCP server "github"

mcp_servers:
  - name: filesystem
    url: http://localhost:3001/mcp

  - name: github
    url: http://localhost:3002/mcp
    headers:
      Authorization: "Bearer ${GITHUB_TOKEN}"

runtime:
  streaming_enabled: true
  resource_profile: small

models:
  providers:
    - anthropic
  model_mapping:
    default: claude-sonnet-4-6

How it works¶

At agent load time, Rune calls initialize + tools/list on each MCP server
Discovered tools are injected into the agent's toolset with the prefix {server-name}/{tool-name}
When the LLM calls a tool like filesystem/read_file, the ToolDispatcher routes it to the MCP server via tools/call
Results are returned to the LLM as tool outputs

Tool naming¶

MCP tools are prefixed with the server name from mcp_servers::

MCP server `name`	MCP tool name	Tool name in Rune
`filesystem`	`read_file`	`filesystem/read_file`
`github`	`create_issue`	`github/create_issue`

Add the prefixed names to toolset: so the LLM knows they're available.

Server URL from environment¶

The MCP server URL can be overridden at runtime via environment variables:

RUNE_MCP_{SERVER_NAME}_URL=http://host:port/mcp

Where {SERVER_NAME} is the uppercased name with hyphens replaced by underscores:

`name` in Runefile	Env var
`filesystem`	`RUNE_MCP_FILESYSTEM_URL`
`my-server`	`RUNE_MCP_MY_SERVER_URL`

Tool descriptor (manual)¶

If you prefer not to use mcp_servers: auto-discovery, you can declare MCP tools manually in tools/:

# tools/filesystem_read.yaml
name: filesystem/read_file
version: 0.1.0
runtime: mcp
mcp_server: filesystem
timeout_ms: 5000

Popular MCP Servers¶

Server	Tools	Run
Filesystem	`read_file`, `write_file`, `list_directory`	`npx @modelcontextprotocol/server-filesystem /path`
GitHub	`create_issue`, `get_pr`, `search_repos`	`npx @modelcontextprotocol/server-github`
Fetch	`fetch`	`npx @modelcontextprotocol/server-fetch`
Brave Search	`brave_web_search`	`npx @modelcontextprotocol/server-brave-search`
Postgres	`query`	`npx @modelcontextprotocol/server-postgres postgresql://...`
Puppeteer	`navigate`, `screenshot`, `click`	`npx @modelcontextprotocol/server-puppeteer`

Start any MCP server and add it to mcp_servers: in your Runefile.