AI that answers from your governed data model — and shows its work
Your agents get a typed, permission-checked map of your business instead of raw tables. Here is exactly what works today, what is half-built, and what is still planned.
Point an AI at raw tables and it guesses which column means revenue, in which currency, for whom. A confident wrong number looks exactly like a right one.
SemanticFed gives it a published model to bind to, and stores the plan behind every answer — so any number can be opened and checked.
Everything below is labelled with where it actually stands in SemanticFed today — shipped, in progress, or on the roadmap.
AI-native by architecture, not bolted on
Four architectural choices that make AI a first-class part of SemanticFed rather than a chat window on the side.
Answers bind to your published model
Agents work from your published entities, metrics and relationships, so a query names fields that really exist — not ones that merely sound right.
Agents get your permissions, not new ones
Every action is checked against the same roles and workspace limits a person faces. An agent is just another signed-in caller, with no private door.
Every answer opens into its plan
Each run keeps which sources it touched, which filters reached them, the join order and estimate versus actual — so a disputed number is checkable.
An API agents can already drive
Everything in the product is reachable through one permission-checked GraphQL API, so agents, the command line and the SDKs do what a person can.
From intent to a governed action
Walk the path a request takes. Select any stage to see what happens there and what backs it.
Stage 1 of 5 · Connect
Register sources and watch their health
Databases, warehouses, lakes and APIs are catalogued, validated and health-checked on a recorded lifecycle.
Each source moves through draft, connected, degraded, disconnected or disabled, and every health check records a verdict, a latency and a message. Today a person reads those signals; the planned triage loop will read the same history.
What the AI in SemanticFed actually does
Filter by delivery status, then open any capability for the detail and what backs it. No capability is listed as shipped without something in the product behind it.
Where the AI shows up
The places AI meets the work in SemanticFed — and how far each one has actually got.
MCP & external agents
ShippedThe one AI surface live in code today: a published Model Context Protocol server that hands your whole governed model to an agent, under your own identity.
- 40 tools for sources, health, models, metrics, relationships, queries, runs and policies
- Two guided recipes: build a model end to end, and author, govern, run and inspect a query
- Open gap: irreversible actions are not confirmation-gated yet — grant access deliberately
See it as a real scenario
Every AI capability above shows up in a concrete SemanticFed story. Open one to read the full walk-through.
Let an AI agent run your data platformThe shipped surface: 40 tools, guided recipes and your live catalogue — under exactly the permissions you already have.Read the story
Make every AI answer openable, not just plausibleEach run stores join order, what was pushed down to which source, and estimate against actual cost.Read the story
Keep AI from becoming a way around your access rulesAgents checked against the same roles and workspace limits as people — plus the two gaps we have not closed.Read the story
Start modelling from a draft instead of a blank pagePlanned: turn a connected source into reviewable draft entities, metrics and relationships. Publishing stays human.Read the story
Catch a degrading source before your analysts doPlanned: health history reconciled into a status change on a cadence, with the source owner notified.Read the story
Get caching suggestions from your own run historyPlanned: hot, slow, repeated queries surfaced with a suggested cache setting and the evidence. You apply it.Read the story
Catch a new PII column the day it appearsPlanned: likely-sensitive fields flagged with a deny policy that stays switched off until a person approves it.Read the storyAI you can actually let near your data
An AI-native product has to be governable. Here is where SemanticFed stands on each control — including the parts still being built.
One place where permission is decided
Every operation declares the action, resource and scope it needs, checked in one single place for people, scripts and agents alike. There is no second path around it.
Workspace isolation on every write
Writes are filtered to your workspace and must match exactly one record, so an identifier leaked from another tenant resolves to nothing. Agent calls included.
Provenance you can open
Every run keeps which sources were touched, which filters reached them, the join order, and estimate against reality — so a disputed number is verified, not argued about.
Nothing becomes AI-only
Modelling, publishing, querying and policy changes are all doable by hand today, and must stay that way as AI lands. Automation is an accelerator, never the only route.
Confirmation before irreversible actions
Stated as a gap: an agent is not yet asked to confirm before deleting an entity or archiving a published model. The mechanism exists; arming it here is tracked work.
Row and column policy when a query runs
Also a gap: policies are written and switched on but not yet applied at run time. Until they are, permission covers who may run something, not which rows come back.
Models
No model provider is wired into SemanticFed today. When the AI layer lands it will use the shared platform AI runtime — provider-switchable, Gemini the intended default — not a bespoke integration here. Bring-your-own-key is not offered and not promised.
- None wired in today — SemanticFed makes no model calls at all
- Gemini — intended default of the shared platform AI runtime, once it ships
- Provider-switchable by design, so no model choice is baked into this product
- Your own agent's model, when it drives SemanticFed over MCP
The honest answers
One AI surface: the Model Context Protocol server. Under it sits shipped machinery — the typed model, cheapest-first planning, real queries against your own sources, and a stored plan per run. No plain-English querying, no in-app assistant, no background loops. Everything else here is labelled roadmap.
Put SemanticFed’s AI to work
Your agents get a typed, permission-checked map of your business instead of raw tables. Here is exactly what works today, what is half-built, and what is still planned.