fuse-browser

Give your AI agent a real, stealth browser. An MCP server + CLI on top of Patchright (stealth Chromium) with a Playwright fallback.

Your agent gets a real browser — per-country identity (locale/currency/timezone/geo), stealth fingerprint, self-healing actions, an indexed snapshot with stable refs (piercing Shadow DOM + iframes), multi-step plans, structured extraction, visual diff, and human guardrails for payments and bookings. It drives real Chromium, so it reads Next.js / SPA pages after hydration — not just static HTML.

49 MCP tools · stealth + rotating proxies · HTTP fast-path (single, batch & crawl) · full-site content + screenshot snapshots · structured per-card product extraction · form fill + file upload · hover + drag&drop · PDF export · cookies / permissions · network mocking · clipboard · virtualized-list scraping + autoscroll · tabs / dialogs / downloads · console + network logs · MCP screenshot resources · FUSE_CAPS tool-group filtering · named auth profiles (cookies + localStorage + IndexedDB, saved at login) · blockResources · HAR record/replay · pixel visual-diff · human handoff + live view.

Install

# Register with Claude Code (or any MCP client) — user scope = all projects
claude mcp add fuse-browser --scope user -- npx -y @fusengine/browser-mcp

// …or add it to your MCP config manually:
{ "mcpServers": {
  "fuse-browser": { "command": "npx", "args": ["-y", "@fusengine/browser-mcp"] }
}}

Chromium installs automatically. That's it — now just ask your agent in plain language:

"Find a hotel in Annemasse this Friday under CHF 100." "What's my Google rank for 'agence web vevey' (CH)?" "Screenshot localhost:3000 in mobile and desktop."

Prefer a terminal? Install the CLI: npm i -g @fusengine/browser-mcp

fuse-browser probe https://example.com --extract-prices
fuse-browser fetch https://books.toscrape.com/ --extract-prices   # no browser, ~10× faster
fuse-browser products "https://www.digitec.ch/en/search?q=macbook" --limit 20   # structured cards → sort to find the cheapest

How it works

An LLM runs a perceive → decide → act loop through the tools: browser_open → browser_navigate → browser_snapshot (indexed refs + form state) → browser_act (click/fill/select/pick/upload/hover/drag, returns a page diff) → browser_wait_for → browser_autoscroll (drain lazy lists) → browser_products / browser_collect / browser_extract / browser_screenshot. Sensitive actions (pay / book / checkout) are blocked unless the agent passes humanApproved.

What you get

• Stealth — Patchright neutralizes the real automation signals; per-country identity + rotating proxy pool.
• Agentic targeting — accessibility-style snapshot with stable refs, self-healing click/fill, multi-step plans.
• Vision (Set-of-Marks) — annotate:true on browser_snapshot/browser_act/browser_screenshot draws numbered badges (= each ref) on the page, so vision models see it and target by ref.
• Sees everything — open Shadow DOM, same/cross-origin iframes, and virtualized/infinite lists (browser_collect, browser_autoscroll to drain lazy-loaded results first).
• Structured extraction — browser_products pulls per-card rows ({title, price, currency, url?}, each price tied to its own title) by detecting repeated card containers — works on Digitec, Booking, Amazon… Sort by price to answer "which is the cheapest?". Layout-agnostic prices: prefix/suffix currency, thousands/decimal markup, CH/EU formats. Also exposed as the CLI products command.
• Full session control — multi-tab (browser_tabs: list/new/select/close popups & OAuth windows), native dialog policy (browser_dialog), captured browser_downloads, plus browser_console / browser_network logs to debug why a page misbehaves.
• Fast-path — browser_fetch impersonates a real Chrome TLS fingerprint for server-rendered HTML, no browser launch — returns clean markdown and optional contacts (extractContacts) at ~HTTP speed. JSON APIs / plain text come back verbatim (no HTML mangling). Opt-in browserFallback auto-renders client-side (SPA/CSR) pages in a real browser when the HTTP response is an empty shell (escalated: true). browser_fetch_batch fetches many URLs in parallel (bounded concurrency, errors isolated per URL). browser_crawl walks a whole site (bounded same-origin BFS, robots-honored) → clean markdown per page. browser_shots_batch captures responsive full-page screenshots of many URLs in parallel (see the design of a whole set of pages at once). browser_collect_batch exhausts the infinite-scroll list of many listing URLs at once (crawl finds the pages, collect drains them). browser_site_shots snapshots a whole site in one call — crawl + screenshot each page, returning content and responsive PNGs per page.
• Data out — multi-currency prices, typed CSS extraction, contact extraction (emails/phones E.164, fastPathFirst cascade), a clean→validate→dedupe→emit pipeline, CSV export, Google SERP rank tracking.
• Ops — persistent sessions, auto crash recovery (a crashed page is recreated in the same context and restored to its last URL between calls), opt-in per-host circuit breaker + bounded probe queue/budget + browser_metrics for mass scraping, live view (watch any session — even headless — in your browser), screenshot://{sessionId}/last MCP resource (read a session's current page as a JPEG on demand), storageState auto-save, named auth profiles (profile), blockResources to skip images/fonts/etc. on batch runs, HAR record/replay, pixel visual_diff, human handoff for login/2FA.
• Context control — FUSE_CAPS registers only the tool groups you need (core/batch/extract/debug/live) for a lighter LLM context, and the batch tools emit MCP progress notifications when the client sends a progressToken.

Documentation

Full reference in docs/:

Installation · CLI · MCP tools (44) · Configuration · Sessions · Extraction · Anti-bot & proxies · Library

Disclaimer

Provided as-is under MIT, no warranty. fuse-browser is a neutral, dual-use automation tool built for responsible automation: sensitive actions (payment, booking, checkout, destructive operations) are gated behind human-approval guardrails, and compliance controls — robots.txt respect (respectRobots), rate limiting, contact-extraction filters — are built in and opt-in, to be configured according to your lawful basis and the target's rules. You alone are responsible for complying with applicable law, target sites' Terms of Service, robots.txt, and data-protection rules (GDPR, nLPD). The opt-in captcha solver and contact extraction are for authorized, lawful use only. See LEGAL.md.

License

MIT