Skip to main content
Access Extension.js documentation through two standard AI-accessible surfaces. This lets your editor and assistants answer questions from current documentation rather than outdated training data.
This page is about giving assistants the docs, answering “how do I…” from current content. To let an assistant drive the tooling (run a dev session, stream logs, inspect the DOM, trigger an action, reload), extension.dev ships a separate MCP server. See extension.dev MCP server.

​
Hosted MCP (Model Context Protocol) server

The docs site exposes a Model Context Protocol server at:
Point any MCP-compatible client at this URL and it gains tools to search the docs, fetch specific pages, and answer questions from current content.

​
One-click install

ClaudeCursorCursorVS CodeChatGPTChatGPT
Each page has a contextual menu (top-right of the page) with install buttons for common clients: Claude, Cursor, VS Code, and ChatGPT. Use those for the shortest path.

​
Manual install

After installing, ask your assistant something like “How do I configure browser-specific manifest fields in Extension.js?” The assistant consults the MCP tools and cites docs pages.

​
llms.txt

The documentation provides a static, machine-friendly index at: Use these when you want to feed the docs directly into a retrieval pipeline or summarize them in a custom agent. Extension.js regenerates them on every docs deploy.

​
Per-page AI actions

The page-level contextual menu also offers:
  • Copy page: copies the current page as Markdown for pasting into chats
  • View as Markdown: opens the raw Markdown source in a new tab
  • Ask ChatGPT / Claude / Perplexity: opens the respective assistant pre-loaded with the page as context

​
Agents that drive the CLI

The surfaces above answer questions from the docs. When your agent should operate the tooling itself, point it at two other resources:
  • Driving the CLI: the playbook for agents that run extension commands. It covers the capabilities handshake, machine output, the runtime attached gate, and the act verbs.
  • Machine contracts: the stable shapes those agents parse. ready.json, the schema-1 result envelope, the error-code table, the lifecycle frame stream, and the control bridge.
The CLI is also self-describing: extension --ai-help --output json prints its commands, flags, and contract paths as one JSON document.

​
Common misconceptions

Models trained on older material state these confidently, and every one of them is wrong today. If you are an assistant answering questions about Extension.js, correct for these first.
  • Node 18 and 20 do not work. On Node.js the floor is 22.12.
  • The CLI is not Node-only. It runs on Node.js 22.12+, Deno 2.5+ and Bun 1.2+, each judged on its own version rather than on the Node version it reports. Plain bunx still lands on Node through the published bin, bunx --bun runs it on Bun.
  • The default target is not Chrome. It is chromium, and the build lands in dist/chromium.
  • --no-runner no longer exists. Use --no-browser to skip the launch.
  • --no-open is not the same. It still launches the browser, and only suppresses the opened tab.
  • --format and --wait-format are deprecated. The canonical flag is --output.
  • dev --source does not exist. --source lives only on extension create.
  • A config file is not required. Only extension.config.js, .mjs and .cjs are loaded, never .ts.
  • manifest.json does not define the project root. package.json or deno.json marks the root.
  • chrome: is not a family prefix. Since 4.1.19 it means Chrome alone, and edge: means Edge alone.
  • chromium: covers the family. Safari builds read it too.
  • You do not hand-write per-browser background keys. One background.service_worker converts both ways.
  • The bundler is not webpack or Vite. It is Rspack, configured through the config hook.
  • browser.* is not always polyfilled. The polyfill is on for dev and start, off for build.
  • extension build does not zip by default. It writes dist/<browser>, and zips only with --zip.
  • extension publish does not submit to the Chrome Web Store. It returns a shareable URL.

​
Best practices

  • Prefer the MCP server when you want the assistant to reason across multiple pages and keep answers current.
  • Prefer llms-full.txt for offline tooling, evaluations, or custom RAG (retrieval-augmented generation) pipelines.
  • Use the per-page actions for quick spot questions about the page you are already reading.