Working with a model

An AI agent uses a Graview app through the same declaration a person does: the same acts, as typed tools, under the same permissions and rules. A declaration is a context graph, and a context graph is the one thing a model reads as well as a person does. Graview has two programming interfaces over one declaration: the surfaces a person uses, and the tools, descriptions and skills a model uses. Neither can drift from the other, because both are derived.

The seat

An intelligence provider in the declaration says what a model may do and how words and pictures reach it. The seat gets the same acts a person has, narrowed by the same policy, and its turns land in the same operation log with an author and an intent — watchable from altitude, and undoable out of order. What comes back from a model is a plan before it is a change: one row per thing, what it was sure about, what goes with what if you decline something, then one batch with one undo.

In the app, a person meets the model where the thing is. Right-click a record for what it picks from the acts, places and panels the app already has; press ⌘K to ask about what is selected, and the answer stands on a card beside it, a change it proposes waiting as a ghost until it is accepted. The host lends the model (ai), and with AI off none of it is drawn.

The todo example's scene with the task Pay the deposit selected and, beside it, an answer card proposing to mark it done, with Accept and Dismiss.The same answer card beside the selected task, in the dark scheme.
The todo example, asked with ⌘K.

Typed tools, derived

Derived affordances and the agent tool surface, generated from the same declarations; graview mcp and graview apply host it against a real store. Every act is a tool with a schema and a sentence attached; the tool surface is generated from the declaration and narrowed by the seat's role, so a model is never offered an act it may not take.

The package: @graview/tools.

What a model can read

graview describe reads the declaration out for something that cannot see: what a blank installation meets and in what order, what is drawn and what falls back, what a seat may do, how a model is reached, what is judged. graview docs writes an llms.txt and an agents.md beside the entry, derived, so they cannot drift.

graview describe <entry> [--as <role>] [--views <module>] [--place <slug|id>]

    --place <slug|id> [--seed <snapshot.json>] [--width <px>] [--id <who>] says what one
    place shows that seat instead: headings, figures, lists and problems.
graview docs <entry> [--out <dir>] [--views <module>]

What a model reads before it writes (31 skills)

Skills that teach an assistant the authoring moves, each ending in a real graview check verdict. Installed into a product by graview skills install . — which graview create runs for you — into .claude/skills and .agents/skills, so Claude Code and Codex find the same instructions.

The studio: a model edits the declaration itself

The declaration itself as a graph: kinds, fields, edges, acts, rules, roles and grants as nodes, edited with ordinary mutations in Graview's own interface, checked before they are applied, migrated, and written back as the files graview create writes.

The package: @graview/studio.