Quick Overview
This announcement and demonstration video from OpenAI features Eric Provencher from Developer Experience presenting WebMCP support for Codex. He explains how adding WebMCP tools to websites improves agent interaction and walks through a live 3D modeling build session.
Key Points
- 1.WebMCP allows web developers to add agent-usable tools directly into their websites so Codex can interact with web applications programmatically.
- 2.Using WebMCP tools speeds up automated web interaction compared to relying on visual clicking through a browser.
- 3.In the Codex Modeling Studio demonstration, Codex used browser-local tools to generate, inspect, and refine a 3D panda model.
- 4.Developers are advised to treat Codex as the primary end user when designing and refining website tools, descriptions, and error feedback mechanisms.
Summary
Eric Provencher from Developer Experience at OpenAI introduces WebMCP support in Codex. While Codex has previously been able to navigate and interact with web pages through standard browser clicking, that process can be relatively slow. WebMCP addresses this limitation by allowing developers to integrate structured tools directly into their websites. Codex can discover these tools, understand what actions are available, and execute tasks within web applications with much greater speed and precision.
To showcase the protocol in practice, Provencher presents a demo application called Codex Modeling Studio. In this demonstration, a prompt instructs Codex to create a seated 3D model of a baby panda. Codex connects to the web studio, which exposes 39 browser-local tools for mesh creation, transformations, spherical UV mapping, texture painting, rendering, and project persistence. Codex configures shapes, captures screenshot previews from distinct camera angles without disrupting the user's viewport, and iterates across multiple sculpting and painting passes.
The demonstration highlights a collaborative workflow where the user and the agent work simultaneously on the same web interface through different interaction channels. The human operator evaluates visual outputs and provides textual guidance, while Codex autonomously calls the underlying studio tools to make real-time modifications to the 3D model geometry and surface paint.
Provencher advises developers creating agent-ready websites to spend time dogfooding their implementations directly with Codex. Because the model itself is the entity executing the tools, developers must ensure tool definitions are discoverable, descriptions are unambiguous, and unnecessary tools are trimmed down. Providing dedicated feedback tools within the application allows Codex to flag runtime errors and indicate when adjustments are needed, establishing a reliable foundation for agent-driven web workflows.
Introducing WebMCP Support for Codex
While Codex has long possessed the ability to navigate websites by clicking through visual interfaces, this browsing method can be slow. OpenAI is releasing WebMCP support in Codex, enabling developers to expose tools directly inside their web applications. Codex can pull these specifications down and immediately understand the site's capabilities to interact with it efficiently.
Demonstration with Codex Modeling Studio
Eric Provencher demonstrates WebMCP using Codex Modeling Studio, an interactive 3D modeling web application. Rather than clicking through manual controls, Codex accesses 39 browser-local tools to sculpt geometry, inspect viewpoints, take screenshots from multiple angles, and iterate on a 3D-printable baby panda model while receiving collaborative instructions from the user.
Best Practices for Building Agent-Ready Tools
Building web tools for an AI model differs from designing interfaces for human users because the AI model functions as the customer. Developers should spend time dogfooding their tools alongside Codex, keeping toolsets concise, refining descriptions for clear discoverability, and providing built-in feedback tools so Codex can report errors or request fixes directly.
The Bottom Line
The video establishes that WebMCP provides a direct programmatic bridge for Codex to discover and use web application tools without relying on visual browser clicks. Through the 3D modeling studio demonstration, it shows how AI agents and human users can collaborate on complex web applications in real time. It leaves the broader developer ecosystem to explore and implement custom WebMCP tools across their own sites.
FAQ
What is WebMCP and how does Codex use it on websites?
WebMCP is a support protocol that lets developers expose programmatic tools directly within their websites. Codex discovers these tools, pulls down their specifications, and executes actions directly on the website instead of clicking slowly through the visual interface.
What tools were demonstrated inside the Codex Modeling Studio web application?
The Codex Modeling Studio demo exposed 39 browser-local tools covering mesh creation, vertex transformations, UV mapping, texture painting, screenshot capture, and project persistence to build a 3D baby panda model.
Why should developers treat Codex as the customer when building WebMCP tools?
Codex is the entity directly calling and discovering the programmatic tools, not the human user. Developers must tailor descriptions, keep toolsets focused, and ensure clear discoverability so the model understands exactly what each tool does.
How can developers assist Codex in reporting errors during tool execution?
Developers can add feedback tools into their web application so Codex can flag errors, report issues, and notify the developer whenever a tool or action requires fixing.
Worth watching for
Web developers and software engineers interested in exposing web application interfaces and programmatic tools to AI agents using WebMCP.
- webmcp
- codex
- openai
- ai-agents
- web-development
- 3d-modeling