Skip to main content
LiteLLM’s mcp_servers config lets you register Computer Use as a tool-providing MCP server and expose it to any model call routed through your proxy. Your LiteLLM still handles LLM traffic; Yambr stays a tool provider.

​
Pointing LiteLLM at managed Yambr

Models stay wherever you’ve configured them in the same LiteLLM file (OpenAI, Anthropic, local vLLM, etc.). Yambr does not publish a chat/completions endpoint — see Access model.

​
Pointing LiteLLM at a self-hosted Computer Use Server

Put this in the mcp_servers section of your LiteLLM config. The {chat_id} and {user_email} templates get filled from the incoming request context.

​
Networking

If LiteLLM and the Computer Use Server run in the same Docker compose network, use the service name (computer-use-server:8081). Across hosts, use the public URL — matching the server’s PUBLIC_BASE_URL for consistency.
If LiteLLM runs in your compose and Computer Use in ours, either:
  • Attach both to the same external Docker network, OR
  • Point LiteLLM at a publicly reachable URL (with MCP_API_KEY set and a reverse-proxy with TLS in front of 8081).

​
Claude Code through LiteLLM

The sandbox’s sub_agent tool runs Claude Code. To route Claude Code’s own API traffic through LiteLLM, set the gateway vars in the Computer Use Server’s .env:
Full recipe (Azure / Bedrock variants, troubleshooting): Claude Code gateway.

​
MCP-servers-for-sub-agent

LiteLLM can act as an MCP gateway for multiple servers. Pass X-MCP-Servers: github,jira,slack on the incoming Computer Use request — the server writes ~/.mcp.json in the sandbox with URLs templated as {ANTHROPIC_BASE_URL}/mcp/{server_name}. Claude Code picks them up automatically. See Sub-agents.