The semantic layer for AI agents

The external world,
resolved for agents.

Facts with sources and time, not pages. Semantics is a continuously updated world model for agents: public sources resolved into one ontology — entities, relations, events, and the evidence behind each.

Early access opens in small batches. We'll only email you about access — reply to any email to be removed.

as of 2026-09-24 09:00Z

semantics · resolveReading sources

01 · How it works

From a page to a fact,
in four steps.

Semantics reads sources on a schedule, extracts what they claim, resolves who each claim is about, and reconciles sources that disagree — before any agent asks.

01 · Observeentity:K9P1

resolvename "K9P1"

2026-09-21 04:12Z registry 200 4.1 KB

2026-09-21 03:55Z filing 200 1.2 MB

2026-09-20 10:03Z news 200 12 KB

2026-09-18 22:40Z website 200 88 KB

02 · Facts, not pages

Web search returns pages.
Agents need facts.

Every time an agent needs outside context, it searches, reads pages, and redoes the same work from scratch. Semantics does that work once and keeps the result.

  1. 01About a resolved entity, not a string
  2. 02Backed by a source you can open
  3. 03When it was observed, and when it was valid
  4. 04Conflicts kept visible, not silently overwritten

03 · One graph

Every lookup lands
in the same graph.

Each entity you resolve is connected to the rest: owners, officers, places, filings. Your first query answers a question. The graph keeps answering after it.

graph · entity:K9P1resolvedconflict

14entities

61assertions

9sources

2events

entity:K9P1 and the entities one hop away.

04 · Changes become events

Not “row updated.”
Something happened.

A license revoked, a role changed, a contract awarded — each change is an event with before and after, its source, and the evidence behind it. Every answer also says how far it sits from reality.

05 · Where agents use it

Context for agents
that act on the world.

Anywhere an agent would otherwise search, read, and guess, it can ask Semantics and get facts it can cite.

Research agents

Answer who, what, and since when about an entity without crawling ten pages and reconciling them in the prompt.

entity:K9P1 → officers, owners, places

Monitoring agents

Query what changed — a role, an address, a license — as events with evidence, instead of diffing pages.

events since 2026-09-01

Verification

Check a claim against every source that made it, with conflicts shown and the observation time attached.

hasOfficer person:B72D → conflict

Enrichment

Resolve a name, domain, or registry ID to one entity and pull its relations into your own records.

k9p1.example → entity:K9P1

06 · Query it

The hard reasoning happens once,
upstream.

Resolving entities and reconciling sources happens before your agent asks. Even a small model just queries the result — over MCP, or with read-only SQL on the published tables.

agent · mcp · semantics

userWho leads entity:K9P1 now, and since when?

toolsemantics · sql

SELECT value, valid_from FROM value_intervals WHERE entity = 'entity:K9P1' AND …

←person:A31C effective 2026-04-01 observed 2026-09-21 source [1] filing

←conflict [2] website still lists person:B72D (observed 2026-09-18)

agentperson:A31C, effective 2026-04-01 according to a filing [1]. One website [2] still lists person:B72D.