What the words mean
Six terms do most of the work in this product. Asked “what is a world?”, an appliance seeded with the rest of this guide could describe how to author one and never say what one is — so this page exists to be the answer.
The wording follows worlds.embabel.com deliberately: the site and the product describing the same thing differently is worse than either being wrong.
A world is a governed, living knowledge graph of your business, derived from the systems you already run and owned by you.
It is the thing your AI acts in. Enterprise AI stalls on integration rather than on model quality, and no model ships knowing your entities, your systems or your permissions — a bag of tools only leaves it improvising every join. A world is the missing piece: typed, governed, queryable, and yours. It runs on your own machine, and one appliance can hold several.
Concretely a world is a directory of declarative configuration — types, views, agents, apps, the realms it installs — plus the graph those things populate and query.
A realm is how your world learns about a system you already run. A folder of declarative YAML in git. Drop one in and the runtime wires up what it declares: types, actions, goals, APIs, MCP servers, scheduled work, saved queries, whole dashboards. Nothing to compile, and nothing you cannot read before you trust it.
The word to hold onto is join. A realm that brings in data nothing else can connect to is a second database you now maintain. A realm sharing an identifier with what your world already knows — an email address, a company name, a repository — makes every existing thing more useful the moment it lands.
Virtual Cypher
Section titled “Virtual Cypher”One MATCH that spans persisted and live data uniformly.
A normal knowledge-graph query reads what is stored in the graph. Virtual Cypher lets one query also reach data that is not in the graph — a HubSpot contact, a GitHub issue, a semantically related email thread — by fetching it on demand the moment the query reaches for it, splicing it in for the life of that one query, and rolling it back afterwards.
The query author writes ordinary Cypher and does not need to know which labels are persisted and which are fetched live. That is the whole point: no mirroring project, no sync to fall behind.
A saved query, referenced by name as if it were a label. The questions worth asking twice become views; the ones worth asking continuously become routines that an agent runs to tell you when the answer changes.
A view composes: because its rows are nodes of a real or virtual type, you can
MATCH a view and keep traversing from it.
A named colleague that does work in your world. An agent has a job, said in one sentence, and a sponsor: the person who answers for it. It holds the routines that do the work.
An agent is always at one of three stages:
| Off duty | nothing it holds runs |
| On duty, observing | its routines run, and every change they would make is held back and reported instead |
| On duty | its routines run and may change things |
It runs the version its sponsor last signed. Editing a routine, updating the realm it came from, or changing a view it relies on shows up on the agent as an unsigned change, and nothing different happens until the sponsor signs. So an update cannot change what an agent does behind the back of the person who answers for it.
A realm can bring agents with it. They arrive off duty, apart from routines that only read, which may already be observing. An agent is adopted by putting it on duty, and whoever does becomes its sponsor. An agent that still needs something, such as a sponsor or a first signature, says so, and is refused on duty until it has it.
Routine
Section titled “Routine”Work an agent does without being asked — a scheduled job, or a reaction to a signal. A short TypeScript program, cron-driven or event-driven, written in the world or shipped by a realm. Every routine belongs to an agent; one that no agent names is gathered into an agent for its realm, or into your own.
Appliance
Section titled “Appliance”The thing you installed. A knowledge graph, an agent runtime, the realms installed into it, and the surfaces you reach it through — the console, the assistant, the MCP endpoint your coding agent connects to. One appliance, one machine, your data on it.