Using Kagent agents via MCP#

Kagent exposes all running agents via a Model Context Protocol (MCP) server embedded in the control plane HTTP server. This allows you to interact with your Kagent agents using any MCP-compliant client at http://<IP_ADDRESS>/mcp.

Setting up#

Connect your MCP client to the Kagent MCP endpoint using the Streamable HTTP transport.

Local Development#

If you are running Kagent locally, you can port-forward the control plane service:

kubectl port-forward -n kagent svc/kagent-controller 8083:8083

Then use http://localhost:8080/mcp as your MCP endpoint. Otherwise, use the IP address of your kagent control plane.

Example: Cursor Configuration#

Add the following to your Cursor MCP settings:

{
  "mcpServers": {
    "kagent-agents": {
      "url": "http://localhost:8083/mcp"
    }
  }
}

Note: Ensure the port matches your local setup (e.g., 8083).

Example: Claude Code#

To add Kagent to Claude Code:

claude mcp add --transport http kagent http://localhost:8083/mcp

Add --scope project to limit the configuration to the current project.

SSE (Server-Sent Events) is currently not supported. You must use Streamable HTTP. Future updates will include stdio support via the CLI.

Using the MCP Server#

The MCP server exposes two core tools:

1. list_agents: Lists all available agents.
2. invoke_agent: Runs a specific agent by name with a given input. Supports sessionID for continuing conversations.

This architecture enables MCP clients (like Cursor or other agents) to discover and orchestrate Kagent agents as "sub-agents," delegating specialized tasks or cluster actions securely.

Tools Overview#

Example Usage#

Here is an example of asking Claude about a Kubernetes cluster:

In this workflow, the client first calls list_agents to discover capabilities, then calls invoke_agent to execute the k8s-agent with the user's query.

Try it out and see what you can do! Feel free to open issues for feedback or suggestions.