Web2MCP

Record a site's API traffic and expose it as WebMCP tools

Web2MCP turns the APIs a website already uses into tools an AI agent can call — automatically, just by browsing. As you use a site, Web2MCP watches the network requests it makes, learns the shape of each endpoint, and registers them as native WebMCP tools through the browser's document.modelContext API. Any in-browser agent (or the DevTools WebMCP panel) can then discover and invoke them. No config files, no schemas to write, no per-endpoint wiring. ━━━━━━━━━━━━━━━━━━━━ HOW IT WORKS ━━━━━━━━━━━━━━━━━━━━ 1. Record — A page-context script observes fetch and XMLHttpRequest calls, capturing every JSON API the site talks to (its own backend and third-party APIs alike), while filtering out static assets and trackers. 2. Aggregate — Requests are deduplicated into endpoint templates (e.g. /users/123 becomes /users/{id}), with JSON Schemas and required fields inferred from what it observes. Everything a site touches is pooled under one domain. 3. Inject — Each endpoint is registered as a WebMCP tool. Calls replay from the page's own context, so cookies and captured auth headers apply automatically. ━━━━━━━━━━━━━━━━━━━━ FEATURES ━━━━━━━━━━━━━━━━━━━━ • Captures everything, including third-party APIs — pooled under the site you're on • URL templating and JSON Schema inference from live traffic • Native WebMCP injection via document.modelContext • One tool pool per website — subdomains and pages share it • Session-parameter auto-fill for opaque tokens an agent can't know (e.g. a page's session crumb) • Toolbar badge showing how many tools are on the page • Popup listing the tools — including the site's own first-party WebMCP tools • Per-site toggle to disable injection whenever you want ━━━━━━━━━━━━━━━━━━━━ REQUIREMENTS ━━━━━━━━━━━━━━━━━━━━ Web2MCP injects tools through the browser's native WebMCP API, so it needs a WebMCP-capable browser: Chrome 149+ launched with --enable-features=WebMCP, or a Chromium build that ships WebMCP. Without it, Web2MCP still records endpoints (recording-only mode) so tools are ready the moment WebMCP becomes available. ━━━━━━━━━━━━━━━━━━━━ PRIVACY ━━━━━━━━━━━━━━━━━━━━ Everything runs and stays on your machine. Recorded endpoints are stored locally in your browser (chrome.storage.local) and never sent anywhere. No accounts, no servers, no analytics. Captured data can include authentication headers for the sites you visit — it never leaves your device, and you can clear a site's data or disable injection at any time from the popup. ━━━━━━━━━━━━━━━━━━━━ HONEST LIMITS ━━━━━━━━━━━━━━━━━━━━ Web2MCP replays the requests a page actually makes. That works well for cookie-authenticated apps, static bearer tokens, and read endpoints across a huge range of sites — especially your own products. It cannot replay requests protected by mechanisms designed to stop exactly that: auth injected by a service worker below the page's fetch, one-time anti-replay tokens (per-request nonces), or client-side request-body signatures that change with every argument. Best for internal tools, dashboards, and your own web apps.