MCP server

MCP server overview

The Equilibrium MCP server exposes claimcheck, the synthetic data generators, and the source registry as tools inside a Model Context Protocol host — so an LLM can call validate_claim or generate_synthetic_sample directly, rather than answering an economic question from training data and hoping it’s right.

Why this exists

An LLM asked “how much did AI raise task exposure in 2025” will produce a fluent, confident-sounding answer whether or not it has a source for the number. The MCP server removes that ambiguity by giving the model a tool that returns the same structured result documented in validating a claim — sources, assumptions, and a confidence band — instead of a plausible sentence. The model still has to decide how to phrase its response, but the underlying number comes from claimcheck, not from the model’s own weights.

Citations and confidence bands are not optional

Every tool response carries evidence and confidence_band fields, whether the calling model asks for them or not. There is no lightweight mode that returns a bare number — a design choice that costs a small amount of response size in exchange for making it structurally difficult for a model host to strip provenance on the way to the user. If your integration only wants the point estimate, you’re welcome to display only that, but the source and band are always in the payload for whoever’s rendering the final response to check.

Running it

There’s no hosted public instance yet — see authentication for the planned difference between a future read-only public instance and a self-hosted instance with write access to your own generator runs. For the full tool catalog, see tools.