Ansvar AI docs
    ansvar.eu →

    MCP basics

    Model Context Protocol (MCP) is an open standard from Anthropic that defines how an AI client talks to an external tool server. The Ansvar gateway is an MCP server; Claude, Cursor, VS Code Copilot, ChatGPT, Gemini, Le Chat and Copilot Studio are MCP clients (each has a setup recipe). Once a client adds the gateway, the agent can discover and call its tools the same way it would call any other tool.

    What MCP defines

    • Tool discovery — a client lists the server's tools, their JSON schemas, and the scopes they require. The gateway filters that list by tier, so a free-tier client never sees a Company-tier tool.
    • Tool invocation — typed parameters in, typed result out, errors carried in a standard envelope. The agent composes a tool call; the server returns the answer.
    • Prompts — server-defined templates the agent surfaces as slash-commands. Several Ansvar workflow families ship them: tender review (tender-review-decompose, tender-review-regulatory-map, tender-review-red-team, tender-review-audit), threat modeling, gap analysis, FRIA, SCA scoring, and the tabletop exercise. Your client lists them alongside the tools, scoped to your tier like the tools themselves, and each workflow prompt drives its workflow through the same tool calls underneath. Two prompts are not workflow drivers and are served from the free tier up: the guided tour (ansvar-tour) and the document-citation prompt.
    • Transport — the gateway uses streamable HTTP at https://gateway.ansvar.eu/mcp. OAuth 2.1 with Dynamic Client Registration secures the channel.

    Why MCP, not REST

    MCP is built around an agent loop: the agent discovers tools at runtime, picks one, calls it, reads the result, and decides what to do next. REST endpoints assume the caller already knows which endpoint to hit. For a fleet of 224 specialist data sources sitting behind a routing layer, MCP's discoverability is the difference between "the agent figures out which jurisdiction to query" and "the developer wires up 224 SDKs by hand."

    Why MCP, not RAG

    RAG retrieves chunks of embedded text from a vector store and hopes the model synthesises them faithfully. The gateway does the opposite: each MCP behind it owns its corpus natively (regulator databases, FTS indexes, structured schemas) and returns the precise article the agent asked for, with the source URL and publisher attached. Hallucination at the retrieval step is the single biggest failure mode in compliance LLM work; MCP cuts it out by replacing the embedding round-trip with a tool call to the regulator's own data.