Harness-Engineered Browser Automation
The MCP server that guides AI agents.

tl;dr — OpenChrome talks directly to Chrome via CDP with zero middleware, reuses your real login sessions, and is the only browser MCP server with harness engineering — 27 intelligent subsystems that guide, protect, and optimize the AI agent at every step.

Imagine 20+ parallel Playwright sessions — but already logged in to everything, invisible to bot detection, and sharing one Chrome process at 300MB. That's OpenChrome.

Search across 20 sites simultaneously. Crawl authenticated dashboards in seconds. Debug production UIs with real user sessions. Connect to OpenClaw and give your AI agent browser superpowers across Telegram, Discord, or any chat platform.

You: oc compare "AirPods Pro" prices across Amazon, eBay, Walmart,
     Best Buy, Target, Costco, B&H, Newegg — find the lowest

AI:  [8 parallel workers, all sites simultaneously]
     Best Buy:  $179 ← lowest (sale)
     Amazon:    $189
     Costco:    $194 (members)
     ...
     Time: 2.8s | All prices from live pages, already logged in.

Traditional browser automation exposes raw APIs. When the AI agent fails, it's on its own — burning tokens guessing, retrying, and wandering. Harness engineering means the tool itself wraps intelligence around those APIs: preventing mistakes, recovering from errors, and guiding the agent toward efficient behavior.

The bottleneck in browser automation isn't the browser — it's the LLM thinking between each step. Every tool call costs 5–15 seconds of inference time. When an AI agent guesses wrong, it doesn't just fail — it spends another 10 seconds thinking about why, then another 10 seconds trying something else.

Playwright agent checking prices on 5 sites:

  Site 1:  launch browser           3s
           navigate                  2s
           ⚡ bot detection          LLM thinks... 12s → retry with UA
           ⚡ CAPTCHA                LLM thinks... 10s → stuck, skip
           navigate to login         2s
           ⚡ no session             LLM thinks... 12s → fill credentials
           2FA prompt               LLM thinks... 10s → stuck
           ...
           finally reaches product   after ~20 LLM calls, ~4 minutes

  × 5 sites, sequential  =  ~100 LLM calls,  ~20 minutes,  ~$2.00

  Actual work: 5 calls.  Wasted on wandering: 95 calls.

OpenChrome eliminates this entirely — your Chrome is already logged in, and the hint engine corrects mistakes before they cascade:

OpenChrome agent checking prices on 5 sites:

  All 5 sites in parallel:
    navigate (already authenticated)     1s
    read prices                          2s
    ⚡ stale ref on one site
      └─ Hint: "Use read_page for fresh refs"    ← no guessing
    read_page → done                     1s

  = ~20 LLM calls,  ~15 seconds,  ~$0.40

The hint engine watches every tool call across 9 categories — error recovery, blocking page detection, composite suggestions, repetition detection, sequence detection, pagination detection, learned patterns, success guidance, and setup hints. When it sees the same error→recovery pattern 3+ times, it promotes it to a permanent rule across sessions via the Pattern Learner.

OpenChrome isn't just a browser API — it's an intelligent harness with 27 subsystems that work together:

npx openchrome-mcp setup

One command. Configures MCP server + auto-approves tool permissions. Restart Claude Code, then say oc.

claude mcp add openchrome -- npx -y openchrome-mcp@latest serve --auto-launch

{
  "servers": {
    "openchrome": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "openchrome-mcp@latest", "serve", "--auto-launch"]
    }
  }
}

{
  "mcpServers": {
    "openchrome": {
      "command": "npx",
      "args": ["-y", "openchrome-mcp@latest", "serve", "--auto-launch"]
    }
  }
}

Parallel monitoring:

oc screenshot AWS billing, GCP console, Stripe, and Datadog — all at once
→ 4 workers, 3.1s, already authenticated everywhere

Multi-account:

oc check orders on personal and business Amazon accounts simultaneously
→ 2 workers, isolated sessions, same site different accounts

Competitive intelligence:

oc compare prices for "AirPods Pro" across Amazon, eBay, Walmart, Best Buy
→ 4 workers, 4 sites, 2.4s, works past bot detection

oc setup                    # Auto-configure
oc serve --auto-launch      # Start server
oc serve --headless-shell   # Headless mode
oc doctor                   # Diagnose issues

read_page supports three output modes:

DOM mode example:

read_page tabId="tab1" mode="dom"

[page_stats] url: https://example.com | title: Example | scroll: 0,0 | viewport: 1920x1080

[142]<input type="search" placeholder="Search..." aria-label="Search"/> ★
[156]<button type="submit"/>Search ★
[289]<a href="/home"/>Home ★
[352]<h1/>Welcome to Example

DOM mode outputs [backendNodeId] as stable identifiers — they persist for the lifetime of the DOM node, unlike ref_N IDs which are cleared on each AX-mode read_page call.

Action tools that accept a ref parameter (form_input, computer, etc.) support three identifier formats:

Backward compatible — existing ref_N workflows work unchanged. DOM mode's backendNodeId eliminates "ref not found" errors caused by stale references.

Headless mode (--headless-shell) doesn't persist cookies across restarts. Enable storage state persistence to maintain authenticated sessions:

oc serve --persist-storage                         # Enable persistence
oc serve --persist-storage --storage-dir ./state    # Custom directory

Cookies and localStorage are saved atomically every 30 seconds and restored on session creation.

OpenChrome includes built-in defenses against Cloudflare Turnstile and similar anti-bot systems, with additional stealth navigation planned for a future release. See Turnstile Guide for details.

Measure token efficiency and parallel performance:

npm run benchmark                                    # Stub mode: AX vs DOM token efficiency (interactive)
npm run benchmark:ci                                 # Stub mode: AX vs DOM with JSON + regression detection
npm run benchmark -- --mode real                     # Real mode: actual MCP server (requires Chrome)
npx ts-node tests/benchmark/run-parallel.ts          # Stub mode: all parallel benchmark categories
npx ts-node tests/benchmark/run-parallel.ts --mode real --category batch-js --runs 1  # Real mode
npx ts-node tests/benchmark/run-parallel.ts --mode real --category realworld --runs 1  # Real-world benchmarks

By default, benchmarks run in stub mode — measuring protocol correctness and tool-call counts with mock responses. Use --mode real to spawn an actual MCP server subprocess and measure real performance (requires Chrome to be available).

Parallel benchmark categories:

OpenChrome works on servers and in CI/CD pipelines without Chrome login. All 46 tools function with unauthenticated Chrome — navigation, scraping, screenshots, form filling, and parallel workflows all work in clean sessions.

# Single flag for optimal server defaults
openchrome serve --server-mode

--server-mode automatically sets:

• Auto-launches Chrome in headless mode
• Skips cookie bridge scanning (~5s faster per page creation)
• Optimal defaults for server environments

OpenChrome is an MCP server — it responds to tool calls, not standalone scripts. Server-side usage requires an MCP client (e.g., Claude API, Claude Code, or a custom MCP client) to drive it:

MCP Client (LLM) → stdio → OpenChrome (--server-mode) → Chrome

For standalone scraping scripts without an LLM, use Playwright or Puppeteer directly.

A production-ready Dockerfile is included in the repository:

docker build -t openchrome .
docker run openchrome

For fine-grained control, use individual flags instead of --server-mode:

openchrome serve \
  --auto-launch \
  --headless-shell \
  --port 9222

OpenChrome has 49 distinct reliability mechanisms across 8 defense layers — ensuring no single failure can hang the MCP server.

┌─────────────────────────────────────────────────────────────┐
│  Layer 7: MCP Gateway                                       │
│  Rate limiter · Tool timeout (120s) · Error recovery hints  │
├─────────────────────────────────────────────────────────────┤
│  Layer 6: Session Management                                │
│  TTL cleanup · Memory pressure · Target reconciliation      │
├─────────────────────────────────────────────────────────────┤
│  Layer 5: Request Queue                                     │
│  Per-session FIFO · Per-item timeout (120s)                 │
├─────────────────────────────────────────────────────────────┤
│  Layer 4: Circuit Breaker                                   │
│  Element (3 fails) · Page (5 fails) · Global (10/5min)     │
├─────────────────────────────────────────────────────────────┤
│  Layer 3: CDP Client                                        │
│  Adaptive heartbeat · Stale target guard · Page defenses    │
├─────────────────────────────────────────────────────────────┤
│  Layer 2: Reconnection Engine                               │
│  Auto-reconnect (5 retries) · Exponential backoff · Cookie  │
│  restore · Sleep/wake detection                             │
├─────────────────────────────────────────────────────────────┤
│  Layer 1: Self-Healing                                      │
│  Chrome watchdog · Tab health monitor · Event loop monitor  │
│  Disk monitor · Health endpoint (/health, /metrics)         │
├─────────────────────────────────────────────────────────────┤
│  Layer 0: Process Lifecycle                                 │
│  Graceful shutdown · Orphan cleanup · Atomic file writes    │
└─────────────────────────────────────────────────────────────┘

32 configurable timeouts cover every operation from CDP commands (15s) to tool execution (120s) to Chrome launch (60s). Every timeout is independently tunable via src/config/defaults.ts.

Finding elements by natural language instead of CSS selectors:

"Submit button" → normalizeQuery → parseQueryForAX → AX Tree Resolution
                                                          │
                                                     match found?
                                                     /         \
                                                   yes          no
                                                    │            │
                                              [AX result]   CSS Fallback
                                                             + Shadow DOM
                                                             + Scoring

• AX-first: Uses Chrome's accessibility tree — framework-agnostic across React, Angular, Vue, Web Components
• Cascading filter: 4-level deterministic priority (exact role+name → role+contains → exact name → partial)
• 3-tier Shadow DOM: Open roots (JS) + closed roots (CDP) + user-agent roots
• Hit detection: After clicking, reports what was actually hit + nearest interactive element
• i18n: Korean role keywords built-in ("버튼" → button, "링크" → link, "드롭다운" → combobox)

git clone https://github.com/shaun0927/openchrome.git
cd openchrome
npm install && npm run build && npm test

MIT