AIControl Docs

Platform & Framework Guide

Any framework whose agent can act as a generic MCP client works with AIControl with zero AIControl-specific code — point the framework's MCP client configuration at AIControl's gateway URL and agent-scoped token, the same way you'd point it at any other MCP server.

This is true by construction, not independently verified per framework. Whether a given framework's MCP client mode actually supports custom auth header injection and has retry/error semantics compatible with this gateway was not checked framework-by-framework. If a framework's MCP client can't send a custom Authorization header, it is not actually zero-code-supported despite meeting the letter of this claim — treat each entry below as "expected to work," not "confirmed to work," until you've tried it against a live AIControl gateway.

All you need for any of the entries below:

  • Gateway URL: http://your-aicontrol-host:8001/mcp/<server_id>/
  • Auth header: Authorization: Bearer <agent-scoped-token> (see Integration for how to issue one)

LangChain / LangGraph (Python)

Both ship MCP client support (langchain-mcp-adapters for LangChain, langgraph's MCP tool loaders for LangGraph). Configure the MCP server entry with AIControl's gateway URL and the Authorization header above instead of your downstream server's real URL.

CrewAI (Python)

CrewAI's MCP integration takes a server URL and headers the same way — point it at the gateway.

Google ADK

Google ADK's MCP toolset takes a server connection params object supporting a custom base URL and headers — point it at the gateway the same way.

Anthropic Claude Agent SDK / OpenAI Agents SDK

Both support MCP servers as a first-class tool source, configured with a URL and headers. Point either at the gateway the same way.

Any other MCP-client-capable framework

If it can point its MCP client at an arbitrary URL with a custom Authorization header, it works. If it can't send a custom header, it cannot reach an agent-scoped gateway endpoint and this pattern doesn't apply — that is the one concrete requirement every framework above is checked against.

For the four named agent-platform integration points (Amazon Bedrock, Azure AI Foundry, Vertex AI / Gemini Enterprise, Microsoft Copilot Studio) — none of which are "just point an MCP client at a URL" in the same way — see Platform Discovery.