Agent tools: WebMCP first
Working with an open Napkin Lab sketch? A capable browser agent should
discover its native tools and call napkin_describe first.
You do not need to ask it to use WebMCP or export the sketch.
Live document, same tab
napkin_describe reads Frames, blocks, notes, flows, and
selection. For supported edits you request,
napkin_apply_operations changes the live canvas in one
undoable batch. Screenshots still help with visual appearance.
This guide is not a connection indicator: the agent must discover
and successfully call the tools in your studio tab.
Older browser or agent?
The studio works without WebMCP. If the browser or agent cannot expose native tools, use normal browser controls, save/open a .napkin file, or use the skill's relay handoff. File-based MCP is not connected to the live tab. Imports, exports, Trace, and viewport controls also use their regular UI or file paths. No polyfill or special install is required to keep sketching.
Quick test
In your browser agent, ask “What Frames are in this sketch?”
The expected first document read is napkin_describe.
The page cannot force an agent's tool choice or provide a capability
that its browser client does not support.
Install or update the portable skill
curl -fsSL https://napkinlab.pages.dev/skill/install.sh | bash
Drops the skill into .agents/skills/napkinlab/ and
.claude/skills/napkinlab/, plus the relay scripts.
The skill prefers WebMCP automatically for open sketches and includes
compatibility fallbacks. Codex metadata allows implicit invocation.
Older installed copies need an update; review local customizations
first because the installer replaces its managed files.