AGENTS &
API!

Here is the rule we build PopCraft by: if the editor can do it, it is a command. Creating a layer, setting a fill, applying a motion preset, exporting a PNG. Each one is a named command with typed parameters, in one registry. The menus call it. The palette calls it. The agent calls it as a tool. The CLI and the REST API call it. Plugins call it too.

That sounds like plumbing, and it is, but it is the plumbing that makes everything else on this page honest. The agent cannot do a weird half-version of a feature, because there is no second version. A script you write against the CLI today runs the same command your designer clicked this morning. When we add a feature, it is written once and every door reaches it. (The CLI and the API take the ones that need no screen; the rest run in the editor.)

The API lives at /api/v1 and is described by an OpenAPI 3.1 document you can feed to any client generator. Read files, nodes, images and versions. Read and write comments. Create files, replace a document, and for targeted edits, post commands: up to 100 run in order and are saved as one edit. A step can use what an earlier one returned, so "$0.id" is the frame step 0 created.

curl -X POST "https://popcraft.app/api/v1/files/$KEY/commands" \
  -H "Authorization: Bearer pop_…" -H "content-type: application/json" -d '{
  "commands": [
    { "command": "node.create", "params": { "type": "FRAME", "name": "Card", "width": 320, "height": 200 } },
    { "command": "node.create", "params": { "type": "TEXT", "text": "Hello", "parentId": "$0.id" } },
    { "command": "paint.setFills", "params": { "ids": ["$0.id"], "color": "#ff0055" } }
  ]
}'

If a step fails, nothing is saved and you get a 422 that says what each step did. "dryRun": true saves nothing either way. Every file response carries a revision; send it back as If-Match and a write only lands if nobody changed the file in between. People editing at the time see the change merge in as if a collaborator made it, and before the first write in 30 minutes the file’s current state is kept as a version.

Plugins are small web bundles that run in a sandbox, read and edit the open file through the same commands, show their own panels and put widgets on the canvas. Share them with your team or publish them to the marketplace.

Claude Code, Codex, Cursor: an AI coding agent in a terminal can make real PopCraft designs with the CLI and nothing else. It works the way our template gallery was made. It writes the whole design as a script, builds it with write_design, and gets back checks: text that fails contrast, text too small for the size, text off the frame, text under the buttons Instagram, TikTok and Shorts draw on top, and text overlapping text.

Then it renders the frames with vision.lookAt, looks at the pictures, fixes the script and builds again in place with --frame. The loop is the whole method, and our public guide, Designing with an AI agent, is the page to hand it. The result is an ordinary .popcraft file a person can open and keep editing.

1. 1

   Install the CLI

   Run npm i -g @popcraft/cli (Node 25 or newer), or brew install getpopcraft/popcraft/popcraft for one executable with no Node.
2. 2

   Make a token

   Account → Personal access tokens. Give it only the scopes it needs. The secret starts with pop_ and is shown once.
3. 3

   Log in

   popcraft login --token pop_… keeps it in your system keychain. popcraft whoami checks it.
4. 4

   Find the command

   popcraft commands lists what runs; popcraft commands paint.setFills shows a command’s flags.
5. 5

   Run it

   On a file on disk, or on cloud:<key>. Add --dry-run to see what would happen first.
6. 6

   Wire it into CI

   Batch steps into one save, keep going past failures if you want, and read --json output in your scripts.