browser_mcp
Enables Claude to navigate, screenshot, read console/network, and crawl links in HTML documentation via a co-located headless Chromium and VSCode webview pane.
README
browser_mcp — VSCode docs browser + hosted MCP
A single VSCode extension that displays auto-generated, self-contained HTML documentation in a real-DOM pane, and hosts an MCP server so Claude (via Claude Code) can develop and test that documentation UI — navigating, screenshotting, reading console/network, and crawling links.
One extension process owns everything:
┌─ VSCode extension host (one process) ──────────────────────────────────────┐
│ • Docs HTTP server → http://127.0.0.1:<ephemeral>/... (static docs) │
│ • Webview pane → <iframe> of the docs URL — REAL, selectable DOM │
│ • Stateless MCP HTTP server → http://127.0.0.1:8765/mcp │
│ • MCP tool handlers ── drive ──► pane (postMessage) + Playwright (IPC) │
│ • Activity-bar view + commands; spawns & owns the sidecar │
└───────────────▲──────────────────────────────────────────────┬─────────────┘
Claude Code │ HTTP (.mcp.json type:"http") │ stdio JSON
(terminal) │ ▼
│ ┌─ Playwright sidecar (child process) ─┐
└───────────────────────┤ headless Chromium; goto/screenshot/ │
│ click; console + network capture │
└──────────────────────────────────────┘
A navigate updates the pane and the Playwright page in the same handler — always in sync, no lockfile/discovery. The pane shows real DOM (humans read/select); Playwright is the AI's surface (screenshots + inspection); both render the same URL.
Why this shape? VSCode 1.126 routes the built-in Simple Browser to a new "integrated browser" that opens a fresh tab per programmatic navigation, and you can't screenshot a VSCode pane. So we render docs in our own webview and use a co-located headless Chromium for pixels.
Requirements
- Node.js ≥ 20 (developed on 25), VSCode ≥ 1.101, Claude Code.
- A Chromium build for
playwright-core. Fetched automatically on first use — the extension runs theplaywright-coreinstaller (shown as a progress notification) the first time a browser is needed, into the shared Playwright cache (~/Library/Caches/ms-playwrighton macOS,~/.cache/ms-playwrighton Linux,%LOCALAPPDATA%\ms-playwrighton Windows). No manual step; subsequent runs reuse it.
Setup
npm install # installs deps (extension workspace)
npm run build # esbuild → extension/dist/extension.js + dist/sidecar/main.js
npm run install:ext # builds + sideloads into ~/.vscode/extensions, carrying playwright-core
Then Developer: Reload Window in VSCode to activate. (Re-run npm run install:ext + reload after editing the extension — the installed copy is a snapshot.)
Alternative install: npm run package:ext → extension/docs-browser-extension.vsix → Extensions view → Install from VSIX (remove any sideloaded copy first). The .vsix bundles playwright-core, so it installs cleanly on another machine; Chromium is still fetched on first use. (Building the .vsix requires the zip CLI on the build machine; installing it does not.)
Use in another project
The extension is installed per user (in ~/.vscode/extensions), so once installed it's available to every VSCode window — you don't reinstall per project. To use it in another project:
- Point it at your docs. Set
docsBrowser.docsDir(workspace setting) to your generated HTML output directory (relative to the workspace root, or absolute). Activity Bar → Docs Browser → the gear, or runDocs Browser: Open Settings. - Register the MCP server. Run
Docs Browser: Create MCP config (.mcp.json)from the Command Palette (or the view title bar). It writes/merges a.mcp.jsonin the workspace root pointing Claude Code at the hosted server, preserving any other servers already listed. - Reload the window, then start Claude Code in that folder (run
/mcpif the tools don't connect immediately).
That's it — no per-project build or install. Note the single-window / fixed-port assumption below if you keep multiple projects open at once.
Connect Claude Code
.mcp.json (project scope) points Claude Code at the extension-hosted server:
{ "mcpServers": { "docs-browser": { "type": "http", "url": "http://127.0.0.1:8765/mcp" } } }
The extension activates on VSCode startup and binds 127.0.0.1:8765 (configurable via docsBrowser.mcpPort). Start Claude Code in this folder and approve the server. If the tools show "failed" (e.g. Claude Code started before the extension was listening), run /mcp to reconnect — the server is stateless, so it recovers across VSCode reloads.
Usage
- Activity Bar → Docs Browser: an "Open Docs Browser" button + a clickable page list + Open/Refresh; when no docs are found yet, a welcome panel links to the docs-directory setting and the MCP-config command.
- Command Palette:
Docs Browser: Open,Docs Browser: Refresh,Docs Browser: Create MCP config (.mcp.json),Docs Browser: Open Settings. - From Claude: navigate/screenshot/etc.; the pane mirrors what Claude does.
MCP tools
| Tool | Purpose |
|---|---|
open_docs / navigate |
Open/go to a page (mirrors to the pane) |
back / forward / reload |
History + reload |
screenshot |
PNG (fullPage, selector, savePath) |
get_page_title |
{title, url} — confirm where you landed |
get_page_text / get_links |
Rendered text / link graph |
list_pages |
HTML files in the docs dir |
read_console |
JS errors/warnings on the current page |
read_network |
Failed / 4xx / 5xx requests (broken assets) |
check_links |
Crawl internal links → broken targets + orphan pages |
click |
Click by selector or text (pane follows) |
pane_state |
Pane URL + preflight status + recent webview events (sync/debug) |
Settings
docsBrowser.docsDir(defaultdocs) — directory of generated HTML to serve.docsBrowser.mcpPort(default8765) — fixed MCP port (strict; fails loudly on conflict).docsBrowser.maxCrawlPages(default500) — page cap forcheck_links; raise for large doc sets (check_linkswarns when the cap is hit). Applied at activation.
Layout
extension/
src/ types.ts (contracts), extension.ts, tools.ts, mcpServer.ts,
paneController.ts, browserClient.ts, docsServer.ts, checkLinks.ts
sidecar/ main.ts # Playwright child process
test/ *.mjs + integration.ts # standalone module + integration tests
build.mjs # esbuild (extension + sidecar)
build-vsix.mjs # package a .vsix (bundle + inject playwright-core)
docs/ # sample docs (replace with your generated output)
.mcp.json # Claude Code MCP registration (http)
scripts/install-extension.sh
Tests
- Module tests:
node extension/test/{sidecar,pane,mcp,docs}.mjs - Full pipeline (headless, real Chromium + MCP HTTP client):
node extension/run-integration.mjs
Notes & caveats
- MCP over HTTP: if Claude Code starts before the extension is listening, reconnect via
/mcp. Stateless transport handles VSCode reloads. - Pane: VSCode find-in-page (Cmd+F) may not reach inside the cross-origin iframe; keyboard focus inside the iframe can swallow some VSCode shortcuts/copy.
- Sidecar: owned by the extension; group-killed on
deactivateso Chromium doesn't orphan. - Multi-window: a second VSCode window contends for the fixed MCP port (strict → fails loudly); single-window assumed.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.