MCP Configuration

The ContextKit MCP server can be configured for any MCP-compatible client. Below are setup instructions for the most common tools.

Claude Code (stdio)

Claude Code uses a .claude/mcp.json file in your project root for MCP server configuration. The stdio transport runs the server as a child process.

.claude/mcp.json

{
  "mcpServers": {
    "contextkit": {
      "command": "npx",
      "args": ["@runcontext/cli", "serve"],
      "cwd": "/path/to/your/project"
    }
  }
}

Tip

Set cwd to the directory containing your contextkit.config.yaml file. If the config is in the project root, you can omit cwd.

HTTP mode

For shared or remote setups, start the server in HTTP mode and point clients at the SSE endpoint.

Start the server:

npx @runcontext/cli serve --http --port 3000

Client configuration (generic):

{
  "mcpServers": {
    "contextkit": {
      "url": "http://localhost:3000/mcp"
    }
  }
}

Cursor

Cursor reads MCP configuration from its settings. Add the ContextKit server under the MCP section.

stdio

.cursor/mcp.json

{
  "mcpServers": {
    "contextkit": {
      "command": "npx",
      "args": ["@runcontext/cli", "serve"],
      "cwd": "/path/to/your/project"
    }
  }
}

HTTP

.cursor/mcp.json

{
  "mcpServers": {
    "contextkit": {
      "url": "http://localhost:3000/mcp"
    }
  }
}

Verifying the connection

Once configured, your AI tool should discover the ContextKit resources and tools automatically. You can verify by asking the agent:

“What MCP resources are available?”

It should list the context:// resources. You can also test a tool call:

“Search my semantic layer for revenue metrics.”

The agent should call context_search and return results from your context directory.

CLI flags

The serve command accepts the following flags:

Flag	Description	Default
--http	Use HTTP transport (default is stdio)	Off
--port <number>	Port for HTTP server	3000
--host <address>	Host address to bind	0.0.0.0
--context-dir <path>	Path to context files	./context