The pairing that does the work is the object tree plus the image export: the model can both traverse the structure and look at what a component renders as. That is what makes design-to-code and design-system questions answerable from the file itself rather than from a screenshot someone pasted.
A server over Penpot's API. It pulls design files, caches them, and lets the model traverse the object tree or render a single component out as a picture, so design questions can be answered against the real file.
- List every project in the account, and the files inside one
- Retrieve a Penpot file by ID and keep it cached for follow-up questions
- Walk the object tree of any object in the file
- Search objects inside a file by name
- Export an object as an image, and read rendered component images back as a resource
- Command-line helpers that draw a tree view of a file or validate one against the schema
Python 3.12+ and a Penpot account. Install with pip install penpot-mcp or run it with uvx. Credentials go in PENPOT_API_URL, PENPOT_USERNAME and PENPOT_PASSWORD - it signs in as your user, so it sees whatever you see. Penpot's cloud site sits behind CloudFlare; if API calls start getting blocked, log in to design.penpot.app in a browser and clear the challenge, then retry.
One command plus a key — pip install penpot-mcp, then supply credentials