AI agents

Queen runs an MCP server for coding agents at https://queenmq.com/mcp. Connect your agent and it starts every session knowing the Queen model and its traps, because the server hands it a primer as the session opens. While it works, the agent asks the server for tested code in your language, checks what it wrote against the known traps, looks up the errors it meets, and learns which Kafka clients work against Queen and with what configuration. There is no account to create and nothing runs on your machine: the server is one HTTPS endpoint that answers questions about Queen.

Install

claude mcp add --transport http queen https://queenmq.com/mcp

claude mcp add --transport http --scope project queen https://queenmq.com/mcp

{
  "mcpServers": {
    "queen": {
      "url": "https://queenmq.com/mcp"
    }
  }
}

{
  "servers": {
    "queen": {
      "type": "http",
      "url": "https://queenmq.com/mcp"
    }
  }
}

[mcp_servers.queen]
url = "https://queenmq.com/mcp"

Any other agent that speaks MCP over Streamable HTTP connects to the same URL, with no headers and no sign-in.

What your agent gets

The server gives your agent six tools, which it calls on its own as it works:

• guide(topic|page) answers from the documentation: the sections that best match a question asked in plain words, or a whole page by its slug.
• example(task, language) returns code for one operation (a push, a consumer, a transaction, a replay) or a whole example app, in JavaScript, Python, Go, PHP, Rust, C++ or plain HTTP. It answers from snippets cut from programs our test suite runs, so the method names are exact, and labels any code that comes from a docs page instead.
• check(language) is the checklist of known traps for code in that language, from the mistakes that lose, stall or duplicate messages down to the ones that only cost speed. With kafka, it covers Kafka clients.
• explain_error(error) says what an error means and what to do about it, for an HTTP status, a code or a transaction reason from a response body, or a Kafka error name or number.
• kafka_client(client, version?) says whether a Kafka client library works against Queen, which versions we tested, the configuration it needs and where it falls short.
• setup(feature, language?) walks through setting up one feature: KV state, timers, stream processing, ephemeral queues, Kafka clients, the Postgres source and sink, or the S3 sink. It answers with the docs for that feature, the calls in your SDK when you name a language, and the traps that apply.

It also offers three prompts, which you run yourself:

• design turns a description of a flow into a Queen model (queues, partition keys, readers, one transaction per step, timers and KV state) before any code exists.
• review goes through the Queen code in your current changes with the trap checklist and the tested API.
• from-kafka brings a Kafka application over, keeping its Kafka client against Queen’s Kafka facade or moving the code to a Queen SDK.

Claude Code lists the prompts as slash commands, /queen:design, /queen:review and /queen:from-kafka, and typing /mcp__queen__design runs the same prompt. In VS Code’s chat they are /queen.design and so on.

Set it up for your team

Commit the configuration and the server comes with the repository. The files are .mcp.json for Claude Code (the one --scope project writes), .cursor/mcp.json for Cursor and .vscode/mcp.json for VS Code, each holding the entry shown above; recent versions of VS Code read .mcp.json too. Everyone who opens the repository in one of those agents is offered the server, and Claude Code asks each person to approve it once. The examples/ folder of the Queen repository carries all three files, so an agent working on the examples gets the server the same way.

When remote MCP servers are blocked

Some companies don’t allow remote MCP servers. The primer the server hands your agent is also a plain file, so you can append it to your project’s AGENTS.md:

curl -s https://queenmq.com/mcp/AGENTS.md >> AGENTS.md

Agents that read AGENTS.md then start every session knowing the Queen model. They don’t get the tools this way, so the file ends with links to what the docs offer instead: every page as markdown, the error reference and the Kafka client guide.

Privacy

A tool call carries a question and nothing else: a topic, a language, an error code, a client name. No tool takes source code as an argument, so your code stays out of the calls, and review runs inside your agent, against the checklist the server sends back. There is no sign-in and no account, and the server keeps no session from one call to the next.