/fuse-browser-usage

name: fuse-browser-usage
description: "Use when about to call any mcp__fuse-browser__* tool. Routes fetch/crawl/SERP vs live browser session vs screenshot capture, and enforces the 4 ZERO-TOLERANCE rules -- read BEFORE the tool call."
keywords:
  - web fetch
  - crawl
  - SERP
  - browser session
  - screenshot
  - webapp testing
  - console errors
  - visual diff
  - core web vitals
user-invocable: false

<objective> This is the canonical doctrine for every `mcp__fuse-browser__*` call: 4 zero-tolerance rules (fast-path first with no browser launch for reading, one session always closed, batch over per-URL loops, deterministic schema-based extraction over manual snapshot parsing), then routing to one of three reference profiles depending on the goal -- research/docs (fetch, crawl, SERP, no pixels), webapp testing (console, network, live interactions), or visual/design (screenshots, responsive + dark mode, regression diffing).

It covers only the raw browser tooling. The full design pipeline (identity -> generate -> audit) is NOT here -- that lives in `fuse-design:design-web` and the `fuse-design:design-review` gate. </objective>

fuse-browser — Efficient Usage

Canonical doctrine for every `mcp__fuse-browser__*` call. Read this skill FIRST, then load the one reference matching your goal. Default engine is **patchright** (stealth identity auto-generated); the Chromium binary comes from `setup.sh`.

4 Rules (ZERO TOLERANCE)

1. **Fast-path FIRST (no browser launch, ~10× faster)** — `browser_fetch` / `browser_fetch_batch` / `browser_crawl` / `browser_serp_batch` to read a page, bulk-fetch, crawl a site, or scrape Google SERP. Open a live session ONLY when you need interaction, JS rendering, or pixels. 2. **One session, always closed** — `browser_open` once → reuse the `sessionId` across `browser_navigate` calls → `browser_close` when done. Never leak sessions; the 2-min TTL is a safety net, not a substitute for closing. 3. **Batch over loops** — `browser_serp_batch` (N queries), `browser_fetch_batch` / `browser_shots_batch` (N URLs), `browser_screenshot { viewports:[...], colorScheme }` (responsive + dark in ONE call). Never write a per-URL loop. 4. **Deterministic extraction** — `browser_extract_schema` with `containerSelector` (card-by-card, correlated fields) over manual `browser_snapshot` parsing.

Routing — pick the profile, load the reference

| Goal | Profile | Reference | |------|---------|-----------| | Read docs, bulk-fetch, crawl, scrape SERP (no pixels) | Research / docs | `references/research-docs.md` | | Test a running webapp (console, network, interactions) | Webapp testing | `references/webapp-testing.md` | | Screenshots, responsive + dark, visual regression | Visual / design | `references/visual-design.md` |

• **No live session for reading** — if you only need text/HTML/data, the answer is always fast-path (profile 1). Never `browser_open` to read.
• **Validation tools** (any profile): `browser_visual_diff` (regression vs baseline), `browser_metrics` (fuse-browser server counters — probes, durations, queue, RSS; not Core Web Vitals — lab CWV come from Lighthouse via the seo plugin's `scripts/check-cwv.ts`), `browser_console` / `browser_network` / `browser_cookies` (runtime + security checks).
• **Full design pipeline** is NOT here — for identity → generate → audit, use `fuse-design:design-web` then the `fuse-design:design-review` gate. This skill only covers the raw browser tooling.