Skip to content

Model Context Protocol (MCP): using the VIDAI Control Plane as the backend

MCP is the plumbing between a client (Claude Desktop, Cursor, Cline, VS Code) and one or more MCP servers that expose tools and resources. The MCP servers don't talk to an LLM themselves — they hand data to the client, and the client passes it into the model as prompt context. That model call is the control-plane-facing part.

Short version: MCP doesn't need any control-plane config on its own. If you've pointed your MCP host (Claude Desktop, Cursor, Cline, etc.) at the control plane using the per-host guide in this section, every model call it makes — including calls where the prompt is enriched by MCP-server data — flows through us.

Per-host guides

  • Cursor — configure Cursor's model endpoint, use MCP servers as you normally would.
  • Cline — same story; Cline's API Provider setting is what matters for the model call.
  • Claude Agent SDK — if you're building a custom MCP host in Python, point the Anthropic client at the control plane.
  • OpenAI Agents SDK — the Agents SDK has MCP client support; point its OpenAI client at the control plane.
  • Claude Desktop: in Claude Desktop's settings, its built-in model is Claude — this call bypasses any proxy. For MCP work you want fully routed, use a code-based MCP host (Cline in VS Code, the Claude Agent SDK, or the OpenAI Agents SDK) with the endpoint pointed at us.

What the control plane sees

Every prompt turn from an MCP host includes whatever context the MCP servers contributed (resource text, tool schemas, tool results). That prompt lands in Request Logs in full. Guardrails apply to the enriched prompt as-is — sensitive data pulled in from an MCP server IS subject to your control plane's guardrail policies.

Attribution

Attribution is determined by the API key on the model call, not by the MCP server. If your Cursor install uses one VIDAI key and your Cline install uses another, the Chargeback tab attributes each MCP-flavoured session correctly per host.

If something's off

Raise an issue at github.com/vidaiUK/vidai-quickstart/issues describing which MCP host you're using and what you're observing (with the model-side endpoint config masked). We'll get it sorted.

Where to go next