Most "AI for data" stories start with a chatbot bolted onto a dashboard. SemanticFed starts somewhere less glamorous and far more useful: a published Model Context Protocol server that turns the entire governed semantic layer into tools an agent can actually call.
The server covers the real domain rather than a demo slice of it. An agent can register a connector, attach a data source and walk it through its connectivity lifecycle, record a health probe, author a semantic model with entities, dimensions, metrics, and typed relationships, publish that model, write and publish a saved federated query, execute it, and read back the run — plus list and toggle access policies. It also ships guided multi-step prompts, so an agent asked to "stand up a model over our warehouse" follows a sequenced recipe instead of improvising, and exposes the federation schema itself as a readable resource so the agent grounds on the real vocabulary.
The important architectural point is what the server does not do. It holds no privileged path. Every tool call travels the same workspace GraphQL gateway as the app, carries the caller's own identity, and is authorized by the same role-based checks that guard a human's click. An agent cannot read a workspace it was not granted, and it cannot perform an action its caller lacks permission for. Agent access is therefore a capability you grant, not a hole you open.
This is the part of SemanticFed's AI story that exists in code today rather than on a roadmap slide — the tool surface, the guided prompts, and the schema resources are published and callable. What is still being hardened is governance of that surface: confirmation gates on the irreversible operations, and a least-privilege review of which tools an agent should be handed at all. Until those land, treat agent access the way you would treat a service account with write permissions.
Pre-launch product. This describes a shipped agent surface over a pre-launch platform, not a measured deployment or a customer result.
This is an MCP server / editor integration story, not a web UI flow. Agents call the semantic layer through the same gateway as the app, but the walkthrough happens in your MCP client, not in the browser, so there is no clickable "Do it yourself" web walkthrough.
Ready to make this your story?
