The AI for AI

What “the AI for AI” means

ColdState is not a language model. It runs no inference at query time and embeds nothing on the way in — that is the whole point. It is infrastructure built for AI to use: a knowledge layer an agent calls when it needs a fact it can ground, cite, and re-verify later.

Most retrieval APIs are designed for a human reading a results page. ColdState’s primitives only really make sense to a machine — a content hash to pin, a snapshot id to reproduce a run, a verify call to fact-check an earlier citation. That is what makes it the AI for AI: tools an agent uses, not a person.

Built for machines to call

Every capability is reachable two ways: a REST API, and a Model Context Protocol (MCP) server that drops straight into Claude and other MCP-compatible assistants. An agent can start with one call — capabilities — to read a machine-readable manifest of every tool, the live knowledge snapshot, and the determinism guarantees, then self-configure from there.

The agent-native toolset

Beyond search, ColdState exposes a set of reproducibility and verification primitives over a knowledge base of 48.4M entries across 35 domains:

• capabilities — a machine-readable manifest of tools, domains, snapshot, and guarantees. Call it first to self-configure.
• resolve — turn a name or alias into the canonical entry id (deterministic: exact-title match wins).
• fetch — retrieve an entry verbatim by stable id, with a content_hash and kb_snapshot.
• cite — a canonical, reproducible citation (id, domain, source, content hash, snapshot) you can re-verify later.
• verify — confirm a previously-cited fact still matches the live knowledge base. The trust primitive for AI-to-AI fact-checking.
• snapshot — the current knowledge-base fingerprint; pin it so the same query returns identical results.
• related — deterministic neighbors via a concept-cluster graph, each flagged cross-domain or not.
• batch_search — many queries in one call against a single consistent snapshot.
• stats / isomorph — coverage and state stats; cross-domain structural analogs (“same structure, different domain”).

Why determinism is the point for agents

An agent built on stochastic retrieval is hard to trust: the same prompt can fetch different context every run, so failures don’t replay and citations rot silently. Deterministic retrieval fixes that — identical query in, identical ranking out (see cold vs hot). An agent can pin a snapshot for an entire run, cite a fact by hash, and later call verify to prove the source hasn’t changed. Reproducible evals, auditable answers, and citations that stay honest.