👑

MCP QUEEN

Connect MCP Queen to your agent stack

One public endpoint gives an agent MCP server discovery and available operational evidence before a separate connection decision.

Universal endpoint

https://mcpqueen.com/mcp

Transport: Streamable HTTP · Authentication: none · Six read-only discovery/evidence tools plus one quarantined feedback tool.

Trust boundary: MCP Queen returns published endpoints and observations. It does not proxy or authorize the server you choose. See the architecture and trust boundaries →

Choose your client

OpenAIClaudeLangChainLlamaIndexCloudflare AgentsHugging Face

OpenAI

Use the hosted MCP endpoint in a compatible OpenAI client or as an MCP tool in the Responses API. The example allowlists only the six read-only discovery and evidence tools.

{
  "type": "mcp",
  "server_label": "mcpqueen",
  "server_url": "https://mcpqueen.com/mcp",
  "allowed_tools": ["search_servers", "search_tools"]
}

Run the Responses API example →

Claude

Claude Code supports the remote HTTP endpoint directly. No local server or Docker image is required.

claude mcp add --transport http   mcpqueen https://mcpqueen.com/mcp

Review every MCP Queen tool →

LangChain

Load the remote tools with MultiServerMCPClient. The repository example calls search_servers directly, so it needs no model key.

client = MultiServerMCPClient({
  "mcpqueen": {
    "transport": "http",
    "url": "https://mcpqueen.com/mcp"
  }
})

Run the LangChain example →

LlamaIndex

Use BasicMCPClient for a direct tool call, or expose the read-only subset to an agent deliberately.

client = BasicMCPClient(
  "https://mcpqueen.com/mcp"
)
result = await client.call_tool(
  "search_servers", {"query": "postgres"}
)

Run the LlamaIndex example →

Cloudflare Agents

Call addMcpServer() from an Agent, then expose only the intended read-only MCP Queen tools to the model.

async onStart() {
  await this.addMcpServer(
    "mcpqueen",
    "https://mcpqueen.com/mcp"
  );
}

Run the Cloudflare Agent example →

Hugging Face

Hugging Face Inference Providers' Responses API accepts a remote MCP server URL and an explicit read-only tool allowlist. Model-provider credentials remain separate from MCP Queen.

{
  "type": "mcp",
  "server_label": "mcpqueen",
  "server_url": "https://mcpqueen.com/mcp",
  "allowed_tools": ["search_servers", "search_tools"]
}

Run the Hugging Face example →

Recommended agent flow

Find → inspect evidence → decide → connect direct.

  1. Call search_servers or search_tools for the task.
  2. Call get_server_grade and get_trust_receipt for candidates.
  3. State which evidence is observed, provisional, or unaudited.
  4. Evaluate the selected server's own permissions, authentication, and data boundary.
  5. Only then configure or authorize that server in a separate step.

Questions developers ask

Does MCP Queen require an API key?

No. The MCP Queen endpoint is public, uses Streamable HTTP, and requires no authentication. An agent framework may still require credentials for its own model provider.

Can an agent connect directly to a server it finds?

Yes. MCP Queen returns the server's published endpoint and available evidence. Connection and authorization happen separately and directly with the selected server.

Does an A operational grade prove that an MCP server is secure?

No. Operational grades measure observed protocol behavior, tool metadata, latency, and provenance. Security, data integrity, citations, claims, response benchmarks, and reviewed field use remain separate evidence dimensions; missing evidence is unaudited.

Need a no-code setup path? Use the separate ChatGPT and Claude connection guide. Runnable framework examples and their dependency instructions live in the MCP Queen repository. See the prepared workflow demo page for the chapter and caption structure.