Weave

Inspiration

Turning a vague idea into a real plan is the tedious part. You know roughly what you want, but breaking it into steps, ordering them, and laying it out takes time. We wanted your browser agent to do that with you, on a canvas, in about a minute.

What it does

Weave is a WebMCP idea-to-plan canvas that runs in the browser. You tell your agent a goal, plan a 3-day Lisbon trip, map a product launch, outline an essay, and it builds a live plan board with you: it drops in cards, links the ones that depend on each other, groups them into columns, lays them on a timeline, and summarises. You drag, edit, and refine while the agent reflows. In under a minute you have a real, usable plan you can export. No account, no data to load; it opens ready to use.

How we built it

Weave is built on WebMCP. The page registers its tools with document.modelContext.registerTool: addCard, updateCard, linkCards, groupCards, setSchedule, reflow, summarizePlan, exportPlan, and clearBoard. The user's in-browser agent (ChatGPT's in-app browser, or Chrome with WebMCP) discovers those tools and calls them with structured arguments; our handlers run in the page and draw the board. The agent is not typing into a box or scraping a screen, it is calling real, typed functions we exposed, so the plan it builds is structured data you can keep and export, not prose. clearBoard is guarded with a confirm so a destructive action never happens silently. Everything is vanilla HTML, CSS, and ES modules, no framework, no build step, no runtime network, so it deploys as a static site and runs inside ChatGPT's browser. A small document.modelContext polyfill lets the same registered tools run from a button when no WebMCP host is present, so it is always demoable.

Challenges we ran into

Designing a tool surface expressive enough to build any plan yet simple enough for an agent to use well. Rendering a smooth, draggable canvas by hand with no library, so the app stays self-contained. Keeping every agent action structured, so the result is real board data you can export rather than a picture of a board.

Accomplishments that we're proud of

A person and an agent co-authoring a real plan on a shared canvas, in under a minute, with zero setup. A genuinely load-bearing use of WebMCP: the agent operates the whole canvas through registered tools. It runs keyless in the browser and is covered by a real test suite.

What we learned

WebMCP turns "the agent describes a plan" into "the agent builds a plan you can use," because the tools produce structured board data, not text. Zero setup is itself a feature: the less a judge has to configure, the more they actually engage with the idea.

What's next for Weave

More layouts, richer export to calendar and docs, and real-time multiplayer so a person and their agent, or a whole team, can plan on the same board together.

Built With

• ai-agents
• browser-agents
• chatgpt
• javascript
• model-context-protocol
• planning
• webmcp