Claude Chrome Parallel (CCP)

Ultrafast parallel browser MCP for Chrome — v1.0.0 released Feb 21, 2026

Automate your actual Chrome — with all your logins, cookies, and sessions intact. Run 20+ parallel browser sessions from Claude Code without logging in to anything, ever again.

"Screenshot my AWS, Stripe, Vercel, GitHub, and Slack dashboards"

Playwright MCP (sequential, login each site):
  AWS    🔐━━━━━━━━━━━━━━ login 45s ━━━━ task
  Stripe              🔐━━━━━━━━━━━━━━ login 40s ━━━━ task
  Vercel                           🔐━━━━━━━━━━━━━━ login 50s ━━━━ task
  GitHub                                        🔐━━━━━━━━━━━━ login 35s ━━━━ task
  Slack                                                      🔐━━━━━━━━━━━━ login 40s ━━━━ task
  Total: ~250s | Memory: ~2.5 GB (5 browser instances)

CCP (parallel, zero auth):
  AWS    ━━━━ 3s ✓
  Stripe ━━━━ 3s ✓
  Vercel ━━━━ 3s ✓
  GitHub ━━━━ 3s ✓
  Slack  ━━━━ 3s ✓
  Total: ~3s  | Memory: ~300 MB (1 Chrome, shared contexts)

  80x faster. 8x less memory. Zero logins.

Why CCP Is Fast

This is not a speed optimization. It's a structural change.

Playwright MCP:  [blank browser] → login → task → close  (repeat per site)
CCP:             [your Chrome]   → task                   (already logged in)

Playwright creates a new browser per site. Each one needs: navigate → type email → type password → solve 2FA → wait for redirect. That's 30-120s per site, and it's sequential. You're spending 95% of the time on authentication, not the actual task.

CCP connects to your existing Chrome via CDP. You're already logged in to everything. Workers run in parallel. The speed advantage compounds with every site:

| Sites | Playwright MCP | CCP | Speedup | |:-----:|:--------------:|:---:|:-------:| | 1 | ~50s (login + task) | ~3s | 17x | | 3 | ~155s (sequential) | ~3s (parallel) | 50x | | 5 | ~250s | ~3s | 80x | | 10 | ~500s | ~3s | 160x |

Memory

Playwright spawns a separate browser process per session (~500MB each). Five sites = 2.5GB.

CCP uses one Chrome with lightweight browser contexts (like incognito windows sharing the same process). Five Workers = ~300MB total. That's 8x less memory — and it stays flat whether you run 5 or 20 Workers.

Bot Detection Immunity

Playwright runs headless browsers with detectable fingerprints. Cloudflare, reCAPTCHA, and anti-bot systems can flag them.

CCP uses your actual Chrome — real fingerprint, real cookies, real browsing history. It's indistinguishable from you clicking around manually, because it literally is your browser.

Core Features

Zero Auth

Your actual Chrome profile. Gmail, Slack, AWS, Stripe — already logged in.

No credentials. No OAuth. No 2FA loops.

20+ Parallel Workers

All Workers run simultaneously in isolated browser contexts.

5 sites in ~3s, not ~250s.

8x Less Memory

One Chrome process, shared contexts. Not N separate browser instances.

5 Workers ≈ 300MB, not 2.5GB.

Worker Isolation

Each Worker gets separate cookies, localStorage, and sessions.

5 accounts on the same site. Simultaneously.

Quick Start

# One command. That's it.
npx claude-chrome-parallel setup

# Restart Claude Code — just say "ccp".

Updates are automatic. The MCP server runs via npx, so you always get the latest version.

{
  "mcpServers": {
    "chrome-parallel": {
      "command": "npx",
      "args": ["-y", "claude-chrome-parallel", "serve", "--auto-launch"]
    }
  }
}

Or: claude mcp add claude-chrome-parallel -- npx -y claude-chrome-parallel serve --auto-launch

npx claude-chrome-parallel setup                  # Global — all projects (default)
npx claude-chrome-parallel setup --scope project  # Project — this directory only

Global writes to ~/.claude.json. Project writes to .mcp.json in the current directory.

Just Say ccp

After setup, ccp is your magic word. No flags, no config, no boilerplate. Just tell Claude what you want with "ccp" and it happens.

You: ccp screenshot my Gmail inbox
Claude: [Already logged in. Screenshot taken. Done.]

You: ccp check my AWS billing and Stripe revenue at the same time
Claude: [2 Workers, parallel, 2.1s — both dashboards captured]

You: use ccp to compare iPhone prices on Amazon, eBay, and Walmart
Claude: [3 Workers, 3 sites, simultaneously]
        Amazon:  $999 | eBay: $945 ← lowest | Walmart: $979

How it works: CCP uses the MCP protocol's native instructions field to teach Claude the keyword automatically. No CLAUDE.md injection. No hooks. Just install and go.

What You Can Do

20-Site Parallel Crawling

You: ccp crawl these 20 competitor sites and extract their pricing

Claude: [20 Workers, 20 sites, simultaneously — all in your logged-in Chrome]
        site-01:  $49/mo  ✓  (1.2s)
        site-02:  $59/mo  ✓  (0.9s)
        ...
        site-20:  $39/mo  ✓  (1.4s)
        Total: 2.8s | Sequential: ~60s | Speedup: 21x

Each Worker runs in an isolated browser context. One Chrome process, 20 parallel sessions, ~300MB total. Not 20 separate browsers eating 10GB of RAM.

Multi-Cloud Dashboard Monitoring

You: ccp screenshot my AWS billing, GCP console, Azure portal, Stripe,
     and Datadog — all at once

Claude: [5 Workers — already logged into every cloud provider]
        aws-billing.png      $12,847/mo  ✓
        gcp-console.png      $8,291/mo   ✓
        azure-portal.png     $3,104/mo   ✓
        stripe-revenue.png   $47,230 MRR ✓
        datadog-metrics.png  99.7% uptime ✓
        Time: 3.1s (not 10+ minutes of login screens)

No OAuth tokens. No service accounts. No API keys to rotate. You're already logged in.

Automated Regression Testing (QA)

# Full regression suite — 10 flows, 10 Workers, one command
claude -p "ccp test these 10 critical flows on staging.myapp.com:
  login, signup, checkout, search, profile-edit,
  password-reset, file-upload, notifications, admin-panel, API-docs"

# Sequential: ~15 minutes.  CCP: ~90 seconds.
# Run this before every deploy.

Each Worker gets an isolated session — no cookie contamination between test flows. Test multi-user scenarios (admin + regular user) simultaneously.

Competitive Intelligence at Scale

You: ccp monitor pricing on Amazon, Walmart, Target, Best Buy, and Costco
     for "Sony WH-1000XM5" — compare and find the lowest

Claude: [5 Workers, 5 retailers, parallel]
        Amazon:    $278
        Walmart:   $298
        Target:    $279
        Best Buy:  $249 ← lowest (sale)
        Costco:    $269 (members only)
        Time: 2.4s | All prices from live pages, not cached APIs

Works on sites with bot detection because it's your real Chrome — real cookies, real fingerprint, real browsing history.

Multi-Account Operations

You: ccp check order status on my personal and business Amazon accounts,
     plus my eBay seller dashboard — all at the same time

Claude: [3 Workers, 3 isolated sessions]
        Amazon Personal:  2 packages arriving tomorrow
        Amazon Business:  Purchase order #4521 approved
        eBay Seller:      3 new orders, $847 revenue today
        Time: 2.1s

Same site, different accounts, simultaneously. Each Worker has its own cookies and session state.

OpenClaw Integration

OpenClaw agent (Telegram/Discord/Signal)
  └─ "Check my AWS billing and Stripe revenue"
     └─ CCP: 2 Workers, parallel, already logged in
        ├── AWS:    $12,847/mo  ✓  (1.8s)
        └── Stripe: $47,230 MRR ✓  (1.5s)
     └─ Response delivered to your chat in ~3s

OpenClaw is an open-source AI agent that runs locally with 50+ integrations across chat, productivity, and automation tools. CCP gives OpenClaw agents authenticated browser superpowers — your agent can browse the web as you, across any site you're already logged into.

• Chat-driven dashboards: "Check my billing" via Telegram → CCP screenshots your AWS/Stripe/Vercel in parallel
• Scheduled monitoring: OpenClaw cron + CCP = automated daily reports from authenticated dashboards
• Multi-platform actions: One chat command triggers parallel browser operations across 10+ sites simultaneously

Orchestration

CCP's core mission is Chrome parallel automation innovation. Simple single-site tasks work great with basic ccp commands. But for complex multi-step workflows, CCP is designed to work with orchestration layers:

| Complexity | Approach | Example | |:----------:|:--------:|:--------| | Simple | ccp command | "ccp screenshot my Gmail" | | Multi-site | workflow_init | "Compare prices across 5 retailers" | | Complex | Orchestrator | Full regression suite, multi-step data pipelines |

Recommended orchestrators:

• oh-my-claudecode — Multi-agent orchestration for Claude Code. Use autopilot, ralph, or ultrawork modes with CCP for persistent, self-correcting browser workflows.
• superpowers — Agent coordination framework. Combines CCP's parallel browser capabilities with structured task pipelines and verification loops.

Comparison

| | Playwright MCP | Browserbase | Chrome Extension | CCP | |---|:---:|:---:|:---:|:---:| | Auth overhead per site | ❌ 30-120s | ❌ 30-120s | ✅ 0s | ✅ 0s | | 5-site authenticated task | ~250s | ~250s + cost | N/A (1 session) | ~3s | | Memory (5 sessions) | ~2.5 GB | N/A (cloud) | N/A | ~300 MB | | Uses your Chrome logins | ❌ Blank browser | ❌ Cloud browser | ✅ | ✅ | | Concurrent sessions | ⚠️ Limited | ✅ (paid) | ❌ 1 (crashes) | ✅ 20+ | | Multi-account isolation | ❌ | ✅ (paid) | ❌ | ✅ | | Runs locally | ✅ | ❌ Cloud only | ✅ | ✅ | | Free | ✅ | ❌ | ✅ | ✅ | | No bot detection | ❌ Headless | ❌ Fingerprinted | ✅ | ✅ | | Device emulation | ✅ | ✅ | ❌ | ✅ | | Network simulation | ✅ | ❌ | ❌ | ✅ | | Workflow orchestration | ❌ | ❌ | ❌ | ✅ | | Adaptive Guidance | ❌ | ❌ | ❌ | ✅ | | Domain Memory | ❌ | ❌ | ❌ | ✅ |

Adaptive Guidance

The biggest time sink in LLM browser automation isn't execution speed — it's wrong tool choices, missed page state, and pointless retries. Each mistake costs 3-10 seconds of LLM inference. Three mistakes and you've wasted 30 seconds before anything useful happens.

CCP injects contextual _hint fields into every tool response to prevent this:

click_element → Error: "ref not found"
  _hint: "Refs expire after page changes. Use read_page for fresh refs."
  → LLM self-corrects. No retry loop. No wasted tokens.

navigate → title contains "Login"
  _hint: "Login page detected. Use fill_form for credentials."
  → LLM skips straight to form filling.

find → computer(click) pattern detected
  _hint: "Use click_element to find+click in one call."
  → Eliminates unnecessary intermediate steps.

21 static rules across 6 priority tiers + an adaptive memory system that learns from your usage. When the same error→recovery pattern appears 3 times, it's promoted to a permanent hint — persisted across sessions in .chrome-parallel/hints/learned-patterns.json.

| Tier | Priority | Examples | |------|----------|---------| | Error Recovery | 100 | Stale refs, tab not found, timeouts, null elements | | Composite Hints | 200 | find+click → click_element, multiple form_input → fill_form | | Repetition Detection | 250 | Same-tool error streaks, A↔B oscillation loops | | Sequence Detection | 300 | Login page detection, navigate→screenshot without wait | | Learned Patterns | 350 | Automatically discovered error→recovery correlations | | Success Hints | 400 | Post-click navigation check, form submission verification |

Memory Architecture

Workers waste 2-3 tool calls per session re-discovering selectors. CCP's domain memory system eliminates this by persisting what works across sessions.

┌──────────────────────────────────────────────────────────┐
│                    Agent (LLM)                           │
│                                                          │
│  "article[data-testid='tweet'] worked → remember this"  │
│                                                          │
│         ┌──────────┐  ┌──────────┐  ┌──────────┐        │
│         │  RECORD   │  │  QUERY   │  │ VALIDATE │        │
│         └────┬─────┘  └────┬─────┘  └────┬─────┘        │
└──────────────┼─────────────┼─────────────┼───────────────┘
               │             │             │
       ┌───────▼─────────────▼─────────────▼───────┐
       │           MCP Server (Store)               │
       │                                            │
       │  ┌─ x.com ────────────────────────────┐   │
       │  │ selector:tweet  → article[data-...] │   │
       │  │ tip:scroll      → scroll first      │   │
       │  │ avoid:read_page → use JS extraction │   │
       │  └─────────────────────────────────────┘   │
       │                                            │
       │  Confidence: +0.1 success / -0.2 failure   │
       │  Pruning: <0.2 removed, 30d stale cleaned  │
       │                                            │
       │  ~/.claude-chrome-parallel/memory/          │
       └────────────────────────────────────────────┘
               │
       ┌───────▼────────────────────────────────────┐
       │        Worker Prompt Injection              │
       │                                             │
       │  workflow_init("x.com") automatically       │
       │  injects domain knowledge into each worker: │
       │                                             │
       │  ## Domain Knowledge (x.com)                │
       │  - selector:tweet: article[data-...]  0.8   │
       │  - tip:scroll: scroll first           0.7   │
       └─────────────────────────────────────────────┘

Three MCP tools — agent decides what to store, server handles persistence:

| Tool | Purpose | Example | |------|---------|---------| | memory_record | Store knowledge after success | {domain: "x.com", key: "selector:tweet", value: "article[data-testid='tweet']"} | | memory_query | Retrieve before site interaction | {domain: "x.com"} → all entries sorted by confidence | | memory_validate | Feedback after using knowledge | {id: "dk-...", success: true} → confidence 0.5 → 0.6 |

Tools (47)

| Tool | Description | |------|-------------| | navigate | Go to URL, back/forward | | computer | Screenshot, click, keyboard, scroll | | read_page | Parse page structure (accessibility tree) | | find | Find elements by natural language | | click_element | Find and click in one step | | wait_and_click | Wait for element, then click | | form_input | Set individual form values | | fill_form | Fill multiple fields + submit in one call | | javascript_tool | Execute JavaScript |

| Tool | Description | |------|-------------| | user_agent | Set User-Agent (chrome, safari, googlebot, etc.) | | geolocation | Override location (seoul, tokyo, new-york, etc.) | | emulate_device | Device emulation (iphone-14, ipad-pro, pixel-7, etc.) | | network | Simulate network conditions (3G, 4G, offline) |

| Tool | Description | |------|-------------| | page_reload | Reload page (optional cache bypass) | | page_content | Get HTML content from page or element | | page_pdf | Generate PDF (A4, Letter, landscape) | | wait_for | Wait for selector, navigation, function, or timeout |

| Tool | Description | |------|-------------| | selector_query | Query elements by CSS selector | | xpath_query | Query elements by XPath expression | | cookies | Get/set/delete browser cookies | | storage | Manage localStorage/sessionStorage | | console_capture | Capture console logs (with type filtering) | | performance_metrics | Performance metrics (FCP, load time, JS heap) | | request_intercept | Intercept/block/log network requests | | drag_drop | Drag and drop by selector or coordinates | | file_upload | Upload files to file input elements | | http_auth | Set HTTP Basic Authentication credentials |

| Tool | Description | |------|-------------| | worker_create | Create isolated browser context | | worker_list | List Workers and their tabs | | worker_update | Update worker progress | | worker_complete | Mark worker as complete | | worker_delete | Delete Worker | | tabs_create_mcp | Create new tab | | tabs_context_mcp | Get tab info | | tabs_close | Close tabs | | workflow_init | Initialize parallel workflow | | workflow_status | Check workflow progress | | workflow_collect | Collect results from all Workers | | workflow_collect_partial | Collect results from completed Workers only | | workflow_cleanup | Clean up workflow resources | | execute_plan | Execute a cached workflow plan | | memory_record | Store domain knowledge (selectors, tips) for reuse | | memory_query | Retrieve learned knowledge for a domain | | memory_validate | Report success/failure to adjust confidence |

| Tool | Description | |------|-------------| | batch_execute | Execute multiple tool calls in a single request | | lightweight_scroll | Scroll without screenshot overhead | | ccp_stop | Gracefully shut down the CCP server |

CLI

ccp setup                         # Auto-configure (global)
ccp setup --scope project         # Auto-configure (project only)
ccp serve --auto-launch           # Start with auto Chrome launch
ccp serve --headless-shell        # Headless mode (15-30% less memory)
ccp serve -p <port>               # Custom debugging port (default: 9222)
ccp doctor                        # Diagnose installation
ccp status                        # View sessions
ccp cleanup                       # Clean up old sessions

ccp requires global install (npm i -g claude-chrome-parallel). All commands also work via npx claude-chrome-parallel <command>.

• Memory — Renderer process limits, JS heap caps, forced GC on tab close
• Screenshots — WebP format (3-5x smaller than PNG)
• Cookie Bridge — 30s TTL cache for auth cookie sharing (~10ms vs 2-6s)
• Find Tool — Batched CDP queries (~100ms vs ~400ms)
• Headless Shell — --headless-shell for 15-30% less memory
• Domain Memory — Persistent selector/strategy knowledge, eliminates 2-3 re-discovery calls per session

Development

git clone https://github.com/shaun0927/claude-chrome-parallel.git
cd claude-chrome-parallel
npm install && npm run build && npm test

Compatibility

CCP is a standard MCP server (stdio JSON-RPC). While optimized for Claude Code, it works with any MCP-compatible client:

{
  "mcpServers": {
    "chrome-parallel": {
      "command": "npx",
      "args": ["-y", "claude-chrome-parallel", "serve", "--auto-launch"]
    }
  }
}

Cursor, Windsurf, Codex CLI, or any editor that supports MCP can use CCP with the config above. See Contributing for multi-client testing status.

License

MIT — LICENSE

Disclaimer: Unofficial community project. Not affiliated with Anthropic. "Claude" is a trademark of Anthropic.