OpenAI names ten WebMCP challenge winners showing what happens when sites expose structured tools to agents
WebMCP is a protocol that lets websites surface capabilities — search, compute, editing, navigation — as machine-readable interfaces. Instead of scraping HTML or guessing at undocumented APIs, an agent receives a mechanical schema of what the site can do and invokes it in a controlled way. OpenAI's challenge asked developers to build applications that pair a human with an agent around those tools. The results span a surprisingly wide range: wedding planning, live Jupyter notebook editing, architectural design, astronomy, CRM workflows, fantasy map-making, flower arrangement, and in-browser 3D scan navigation.
Architecture and design: geometry that corrects itself
Alza, from Elioz 404, takes a floor-plan screenshot and turns it into an editable 2D and 3D model. A WebMCP agent checks the geometry in real time as the user moves walls or adds windows; if the dimensions don't add up, the agent flags the issue and suggests a fix. ArchMorph, from xmusfk, goes a step further: the same live building model serves both for collaborative editing with the agent and for an instant 3D walkthrough. Both projects show how a single structured tool — geometry validation — shifts the design loop from trial-and-error to immediate feedback.
Planning with constraints: from weddings to stars
Aisle, by Eri Alipaj, tackles the wedding-seating puzzle. A WebMCP agent receives a guest list, relationships ("the aunt doesn't sit with the ex-husband"), and locked seats, then returns an arrangement that respects every constraint. Roque Nights, by Sebas Media Prod, does something similar for amateur astronomers: the agent compares nights, identifies visible celestial objects for a given location and time, and proposes an observation plan the user can approve or tweak. Mandate, by Harzer Heribert, brings the idea to CRM: the user selects records, fields, and a time limit, and the agent gets a scoped toolset that lets it act only within those boundaries — an approach that reduces the risk of destructive actions.
Creation, research, and education: an agent inside the environment
Observatory, by Martin Hooijmans, lets the user and agent co-create fantasy maps — terrain, rooms, labels — with everything remaining editable through WebMCP. Bouquet Studio, by Jane Chao and Luis Cesar Morales, translates natural-language prompts like "warmer" or "less formal" into a visual flower arrangement the user can still move by hand. Faraday, by TheGreatAxios, lets an agent navigate 3D scans that stay in the browser — no cloud upload, no heavy file transfers — aimed at research and teaching. JupyterLite WebMCP, by Allison Coleman and Juan Mendoza, drops an agent straight into a live notebook: editing cells, running code, reviewing changes, all in the same context the developer is already working in.
What changes next
The common thread across all ten isn't "generic AI" but the fact that the site exposes well-defined capabilities and the agent operates through them instead of guessing. That's a fundamentally different model from RAG or function-calling against a closed API: the tool belongs to the site, the agent is just a consumer, and the human stays in the loop with the ability to approve, modify, or cancel at any stage. OpenAI didn't publish benchmarks to compare the projects, and the challenge is more a technology demonstration than a measured competition. Still, the list signals a direction: as more sites adopt WebMCP or similar protocols, the line between "an app with AI" and "a shared workspace with an agent" will blur, and the developers who understand how to expose the right tools will be the ones shaping what that interaction looks like.