Overview of 979: WebMCP: New Standard to Expose Your Apps to AI
Wes Bos and Scott Tolinski explain WebMCP — a new web specification that lets websites directly advertise and expose their app functionality as "tools" an AI can call. Instead of relying on external MCP servers, brittle browser automation (Playwright/screenshots), or embedding UI widgets, WebMCP lets your site publish the actions it can perform so an AI (or agent) can discover and call them quickly and reliably.
What problem WebMCP solves
- Current AI-to-web interactions are slow and error-prone:
- Agents parse DOM, accessibility trees, or screenshots and then simulate clicks — latency and token costs are high.
- Publishing a separate MCP server duplicates logic and requires extra infrastructure.
- WebMCP lets the website itself be the interface for AI: publish available tools and their schemas alongside your HTML/JS so an AI can call them directly.
How WebMCP works (high level)
- A site publishes "tools" (actions) it supports, with descriptions and input/output schemas.
- Two declaration methods:
- Imperative JS API (example): window.navigator.modelContext.registerTool(...) — register tools, inputs, outputs programmatically.
- Declarative HTML forms: add simple attributes to forms/inputs (tool name, description, param title/description). The browser can infer schema from form fields.
- An AI-enabled agent (or browser extension) visits the site, discovers the tools, and uses natural-language prompts to decide which tool(s) to call and with what data.
Demo recap (grocery app)
- Wes built a simple grocery/kanban app with common actions: add store, add item, rename, move item, check off, get items by store, etc.
- Chrome extension demo:
- The extension discovers site tools and connects to Gemini Flash 2.5.
- Example commands: "Add bananas to Costco shopping list", "Move bananas from Costco to Whole Foods", "Add all items for chicken noodle soup to Whole Foods", "Mark off anything with chicken".
- Results: AI parsed natural-language requests and performed multi-step actions quickly (~5s for some actions).
- Key demo takeaways: fast, tolerant of typos, and can expand shorthand requests into multiple API calls (e.g., ingredient list for a recipe).
Main benefits
- Speed: explicit tool calls are much faster than UI automation.
- Token efficiency: sending compact tool calls + schemas uses far fewer tokens than passing large DOMs/screenshots.
- Mixed UX: users can still visit the site UI while AI can operate via natural language — good for experiences where the full site is preferable (e.g., booking, upsells).
- Low infrastructure overhead: publish tools as part of your website — no separate MCP server required.
- Framework-friendly: frameworks already have schemas/validation/forms — adding a step to publish those as WebMCP tools is straightforward.
Questions, limitations & open concerns
- Early spec: WebMCP is newly released and evolving (Wes was in early access).
- Cross-app workflows: plausible (agents could visit multiple sites and move data between them), but details/UX are still TBD.
- Auth, permissions, and security: not deeply covered in the episode — important topics to expect as the spec matures (how agents authenticate, scope tool access, rate limiting, consent).
- Platform/commerce behavior: large platforms historically limited API access; some companies may resist exposing functionality that bypasses their UI/monetization.
- Not a universal replacement: some sites will still prefer richer in-site UIs, or not expose tools at all.
Use cases & where it fits best
- CRUD-style apps (to-do lists, grocery lists, habit trackers, bookkeeping) where operations map cleanly to tools.
- Productivity workflows where a user has natural-language intents spanning multiple operations (e.g., "add all veggies from last week's list").
- Developers who want AI integration without building/maintaining a separate MCP server or complex automation scripts.
Practical next steps / recommendations
- If you maintain an app with form-based actions, evaluate adding WebMCP declarations in HTML or via the JS API to make it AI-friendly.
- For early experimentation:
- Sign up for the WebMCP early preview (link in episode description).
- Try the Chrome extension demo to see tool discovery in action.
- Watch for evolving guidance on auth/consent before exposing sensitive actions.
- Framework authors: consider adding built-in WebMCP publishing to make adoption easy for apps built on your stack.
Notable quotes / soundbites
- "This is a great way for the web to adapt to AI."
- "It's much more token efficient."
- "The form-based approach is absolutely genius."
Bottom line
WebMCP provides a lightweight, web-native path to let AIs interact with websites by advertising structured tools directly on the site. It promises faster, cheaper, and more robust AI interactions for many app types without extra servers — but the spec is early and questions remain around cross-site flows, auth, and how platform owners will respond. If you run an app with clear user actions, WebMCP is worth exploring now.