Skip to content
PuniDocs
Open Puni

Agents


Connect Claude, Codex, Cursor or ChatGPT to your library, and understand what an agent can and cannot see.

The agent loop: inspect, draft, apply, look, committhe saved revision keeps a checkpoint; in the studio it is one undo step01Inspecta savedrevision02Draftprivate,staged03Applycommands,atomically04Looknumbers orpixels05Commitone revisiona human save in between refuses itevaluate: numbersrender: pixelsThe agent loop: inspect, draft, apply, look, commit01Inspecta saved revision02Draftprivate, staged03Applycommands, atomically04Looknumbers or pixels05Commitone revisionevaluaterendera human save in between refuses itthe saved revision keeps acheckpoint
Inspect, draft, apply, look, commit. Numbers are not a picture; a human save in between refuses the commit.

Guides

An agent edits the same scene you do, through the same commands, and every edit stays undoable. Connection options include a runner on your machine, a token you paste into a coding client, a sign-in for ChatGPT and Claude, and the browser tools a WebMCP browser gets on puni.design.

#On your own machine

Connected runners execute turns on your own machine, through the command line (a preview for Puni's own team). A disconnected runner yields PUNI_RUNNER_REQUIRED; Puni never executes the turn on a deployment model key.

#From your own client

Connect Puni covers ChatGPT, Claude, Cursor, Grok, Grok Bot and coding clients, with a first scene and picture to check the connection. Scopes and tokens says what a token allows and how to revoke one. Prompts has the prompts that work, with the working loop written in.

#What an agent sees

How agents see is the honest page: agents inspect numbers, and a browser without WebGL shows them nothing. Plugins and skills says what is planned.

Read Connect Puni first to use your preferred client with Puni.