@silbercue/chrome

SilbercueChrome

The fastest, most token-efficient MCP server for Chrome browser automation. Direct CDP, a11y-tree refs, multi-tab ready. 24/24 on the hardest benchmark at 20s scripted, beating Playwright MCP and claude-in-chrome.

Built for Claude Code, Cursor, and any MCP-compatible client.

Looking for an alternative to Playwright MCP, Browser MCP, or claude-in-chrome? SilbercueChrome talks to Chrome directly via the DevTools Protocol — no Playwright dependency, no Chrome extension bridge, no single-tab limit. One command to install, zero config, and the best benchmark score in the category. See comparison below.

Why SilbercueChrome?

Every Chrome MCP server has the same problem: too many tokens, too few reliable refs. Screenshots eat 10-30x more tokens than text trees. Selector-based refs break the second the DOM rerenders. Extension bridges (Browser MCP) get stuck on the connected tab. Playwright wrappers spin up a new browser instance for every session.

SilbercueChrome fixes this. It talks directly to Chrome via CDP (same protocol Playwright and Puppeteer use internally), returns an accessibility-tree-based reference map, and caches it across calls so click(ref: 'e5') and type(ref: 'e7', ...) survive scrolls and DOM updates.

What you get	Playwright MCP	Browser MCP	claude-in-chrome	browser-use	SilbercueChrome
Hardest benchmark (24 tests, LLM-driven)	24/24 (~570s)	cannot finish	24/24 (1140s)	17/24 (~1049s)	24/24 Pro: 555s · Free: 755-900s
Scripted benchmark (24 tests)	—	—	—	—	24/24 in ~20s
Multi-tab support	Yes	No (single tab)	Yes	Partial	Yes
Connection	New browser	Extension bridge	Extension	Subprocess	Direct CDP (pipe or WebSocket)
Ref system	Playwright refs	Playwright refs	CSS selectors	Screenshots	A11y-tree refs (stable across DOM changes)
Read page	Screenshot + DOM	Snapshot	DOM dump	Screenshot-heavy	read_page — 10-30x fewer tokens
Drag & drop	Yes	No	Partial	No	Yes (native CDP mouse events)
Shadow DOM + iframe	Yes	Yes	Partial	No	Yes (with OOPIF session support)
Keyboard shortcuts	Yes	Yes	Partial	No	Yes (press_key with real CDP keyboard events)
localStorage/cookies	Yes	No	Partial	No	Yes (via evaluate)
Multi-step plan execution	—	—	—	—	run_plan — server-side plan executor with variables, conditions, suspend/resume
Zero-config install	Yes	Yes	Built-in	Yes	Yes (one claude mcp add line)

Where SilbercueChrome really shines

24/24 on the hardest benchmark in 20 seconds (scripted), 555s LLM-driven — beats every alternative in the category

The test-hardest suite covers 24 patterns that break most browser MCPs: infinite scroll, Shadow DOM, nested iframes, drag & drop, canvas clicks, keyboard shortcuts, contenteditable, async timing races, 10K-element DOMs, localStorage chains, mutation observers, modal form chains. SilbercueChrome Pro clears all 24 in ~555s LLM-driven (scripted: ~21s). Free tier runs 20s scripted / 755-900s LLM. Playwright MCP takes ~570s LLM. claude-in-chrome takes 1140s. browser-use fails 7 tests architecturally.

read_page with a11y-tree refs — 10-30x cheaper than screenshots

Instead of pushing an 800KB screenshot every turn, read_page returns the accessibility tree with stable e5-style refs. Agents read text, find elements, and chain click(ref: 'e5') / type(ref: 'e7', ...) — all in 30KB responses. Screenshots stay available for visual verification via screenshot, but the LLM stops defaulting to them for element discovery.

True multi-tab — virtual_desk, switch_tab, parallel tabs in run_plan

Browser MCP binds to a single "connected" tab via its Chrome extension — cross-tab operations are architecturally impossible. SilbercueChrome uses CDP Target API to enumerate, open, close, and switch between tabs. virtual_desk lists every open tab with stable IDs. switch_tab moves between them without touching the user's active tab. run_plan even supports parallel tab execution.

fill_form — one call for a complete form

Other MCPs make you emit N type calls for an N-field form. fill_form takes a single fields[] array with refs and values, handles text inputs, <select> (by value or label), checkboxes, and radios in one CDP round-trip, and reports per-field status.

observe — watch DOM changes without writing JavaScript

Two modes: collect (watch for N ms, return every text/attribute change) and until (wait for a condition, then auto-click). Use click_first to trigger the action that causes changes — the observer is set up before the click, so nothing is missed. Replaces the typical setInterval/MutationObserver/evaluate dance.

run_plan — server-side multi-step automation

Execute a sequence of tool steps server-side with variables ($varName), conditions (if), saveAs, error strategies (abort/continue/screenshot), and suspend/resume for long-running workflows. Parallel tab execution is a Pro feature.

Quick Start

Install in Claude Code

One command — installs globally for all projects:

claude mcp add --scope user silbercuechrome npx -y @silbercue/chrome@latest

Restart Claude Code. First tool call auto-launches Chrome visible (no headless, no port setup). Done.

Install in Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "silbercuechrome": {
      "command": "npx",
      "args": ["-y", "@silbercue/chrome@latest"]
    }
  }
}

Install in other MCP clients

Any client that supports stdio MCP servers: npx -y @silbercue/chrome@latest with no arguments.

Uninstall

claude mcp remove --scope user silbercuechrome

Free vs Pro

The Free tier gives you 18 tools covering 24/24 benchmark tests in the scripted runner. Pro adds virtual_desk, switch_tab, dom_snapshot, and advanced run_plan features (parallel tabs, operator hooks, ambient context) plus faster internals.

	Free	Pro
Tools	18	21+
Page understanding	read_page	read_page + dom_snapshot (spatial queries)
Tab management	navigate, tab_status	+ virtual_desk, switch_tab, parallel tabs in run_plan
Interaction	click, type, fill_form, press_key, scroll, file_upload, handle_dialog	Same
Observation	screenshot, wait_for, observe, console_logs, network_monitor	Same + ambient page context hooks
Scripting	run_plan (sequential)	run_plan (sequential + parallel + operator hooks)
Last resort	evaluate	evaluate + anti-pattern scanner hints
Benchmark score	24/24	24/24
Benchmark time (scripted)	~20s	~21s
Benchmark time (LLM-driven)	755-900s	~555s

Pro costs $19 USD one-time. Get a license on Polar.sh, then activate via the built-in license command or SILBERCUECHROME_LICENSE_KEY=SC-PRO-... env var.

Tools

Reading & Observation

Tool	Description
read_page	Accessibility tree with stable e-refs — primary way to understand the page. 10-30x cheaper than screenshots. Filter by interactive (default) or all (include static text).
screenshot	WebP capture, max 800px, <100KB. Use for visual verification only — you cannot use screenshots to drive click/type, refs come from read_page.
console_logs	Retrieve browser console output with level/pattern filters
network_monitor	Start/stop/query network requests with filtering
observe	Watch DOM changes: collect (buffer changes over time) or until (wait for condition, then auto-click)
wait_for	Wait for element visible, network idle, or JS expression true
tab_status	Active tab's cached URL/title/ready/errors — mid-workflow sanity check

Interaction

Tool	Description
click	Real CDP mouse events (mouseMoved/Pressed/Released). Click by ref, selector, text, or x+y coordinates. Response includes DOM diff (NEW/REMOVED/CHANGED).
type	Type into an input by ref/selector
fill_form	Fill a complete form in one call — text, <select>, checkbox, radio. Per-field status, partial errors don't abort.
press_key	Real CDP keyboard events — Enter, Escape, Tab, arrows, shortcuts (Ctrl+K, etc.)
scroll	Scroll page, element into view, or inside a specific container (sidebar, modal body)
file_upload	Upload file(s) to an <input type="file">
handle_dialog	Configure alert/confirm/prompt handling before triggering actions

Navigation

Tool	Description
navigate	Load a URL in the active tab. Waits for settle. First call per session is auto-redirected to virtual_desk to prevent blindly overwriting the user's tab.

Scripting

Tool	Description
run_plan	Execute a multi-step plan server-side. Variables ($varName), conditions (if), saveAs, error strategies (abort/continue/screenshot), suspend/resume. Parallel tabs require Pro.
configure_session	View/set session defaults (tab, timeout) and accept auto-promote suggestions
evaluate	Execute JS in the page context. Use for COMPUTE or side effects no tool covers — not for element discovery (use read_page instead). Anti-pattern scanner warns when you reach for querySelector or .click().

Pro tier (additional)

Tool	Description
virtual_desk	Lists all tabs with stable IDs. Call first in every session.
switch_tab	Open, switch to, or close tabs by ID from virtual_desk
dom_snapshot	Bounding boxes, computed styles, paint order, colors. For spatial questions read_page cannot answer.

Benchmarks

Measured on https://mcp-test.second-truth.com — 24 tests in 4 levels (Basics, Intermediate, Advanced, Hardest). Each run is independent, values on the benchmark page are randomized per page-load, and all runs started in a fresh Claude Code session out of /tmp (no project context bias).

MCP	Passed	Time (LLM)	Time (scripted)
SilbercueChrome Pro	24/24	555s	21s
SilbercueChrome Free	24/24	755-900s	20s
Playwright MCP	24/24	~570s	—
claude-in-chrome	24/24	1140s	—
browser-use	17/24	~1049s	—
Browser MCP	—	could not complete	—

browser-use fails 7 tests architecturally: infinite scroll (no container-internal scrolling), drag & drop, canvas click, keyboard shortcuts, contenteditable bold, localStorage+cookie chain, mutation observer. Browser MCP's single-tab extension bridge cannot complete the Tab switch, read, return pattern and becomes unstable over longer runs. See test-hardest/BENCHMARK-PROTOCOL.md for the full protocol and raw JSON runs.

Architecture

SilbercueChrome (Node.js MCP server, @silbercue/chrome)
├── @modelcontextprotocol/sdk (stdio transport)
├── CDP Client
│   ├── WebSocket transport (existing Chrome on :9222)
│   └── Pipe transport (auto-launched Chrome with --remote-debugging-pipe)
├── Auto-Launch: Chrome + optimal flags, visible by default
├── A11y-tree cache + Selector cache
├── Session Manager (OOPIF support for iframes and Shadow DOM)
├── Tab State Cache (URL/title/ready across tabs)
└── 18 Free-tier tools + 3+ Pro-tier tools
    Reading · Interaction · Navigation · Scripting · Observation

Connection priority:

• Auto-Launch (default, zero-config) — starts Chrome as a child process via --remote-debugging-pipe, visible as a window, with all flags set for reliable screenshots and keyboard focus.
• WebSocket (optional) — if you already run Chrome with --remote-debugging-port=9222, SilbercueChrome connects to that instead. Use this to control your own browser with its extensions and login sessions.

Requirements

• Node.js >= 18
• Google Chrome, Chromium, or any Chromium-based browser (auto-detected on macOS/Linux/Windows; override with CHROME_PATH)

Environment Variables

Variable	Values	Default	Description
SILBERCUE_CHROME_AUTO_LAUNCH	true / false	true	Auto-launch Chrome if no running instance found
SILBERCUE_CHROME_HEADLESS	true / false	false	Opt-in headless mode for CI/server environments
SILBERCUE_CHROME_PROFILE	path	—	Chrome user profile directory (auto-launch only)
CHROME_PATH	path	—	Path to Chrome binary (overrides auto-detection)
SILBERCUECHROME_LICENSE_KEY	license key	—	Pro license key (e.g. SC-PRO-...)

License

The core server and all 18 Free-tier tools are MIT licensed — see LICENSE. Use them however you want, commercially or otherwise.

Pro tools (3+ gated tools, parallel tab execution, ambient context, operator hooks, faster internals) require a paid license. The license validation code is in the separate private Pro repository.

Contributing

Issues and pull requests welcome at github.com/Silbercue/silbercuechrome.

Privacy

SilbercueChrome runs entirely on your machine. All browser automation happens locally via CDP. No telemetry, no remote calls, no data sent to any third party.