🚀 antigravity-with-chatgpt

Zero-Dependency Dual-Brain Reasoning & Verification Architecture for Google Antigravity 2.0 & Gemini

Harness the full reasoning power of ChatGPT Web (GPT-4o, o1, o3, Canvas) as an independent cloud brain for your local Antigravity 2.0 IDE.

English | 中文说明

📖 Background & Design Philosophy

When developing with modern agentic coding assistants, local agents like Google Antigravity 2.0 IDE, Gemini CLI, and Claude Code have deep local authority (file I/O, terminal execution, running tests, Git version control). However, when tackling complex architectural refactoring, algorithmic derivation, subtle root-cause debugging, or rigorous code reviews, a single model within a shared context often suffers from confirmation bias and self-rationalizing hallucinations.

Inspired by the notable community project XiaoDuoYa/codex-with-chatgpt ("ChatGPT thinks. Codex works.") and shaped by ChatGPT's rigorous 38-point architectural code review, antigravity-with-chatgpt establishes a robust Proxy-Pull Context Broker pattern:

• Local Agent Owns Execution: File edits, shell terminal commands, build pipelines, and unit tests strictly belong to Antigravity.

• ChatGPT Owns Reasoning & Review: High-level task planning, mathematical derivations, and adversarial code reviews are delegated to ChatGPT Web.

• Zero API Key Consumption: Reuses your existing authenticated ChatGPT Plus/Pro web session in a dedicated Chrome profile—unlimited access to GPT-4o, o1, and o3 without API costs.

• Zero npm Dependencies: Written entirely in native Node.js 22+ standard library (native WebSocket, fetch, crypto, child_process)—lightning fast startup, zero supply-chain attack surface.

Related MCP server: cdp-tools-mcp

🌟 Key Features

• ⚡ Sub-Millisecond DOM Injection: Utilizes execCommand('insertText') combined with Base64 encoding to bypass ProseMirror per-character event overhead, injecting 10,000+ characters in under 5ms.

• 🛡️ Production-Grade Security Boundary:

  • path_guard.mjs: Platform-aware relative path containment, strictly preventing ../ traversal, NUL byte injection, and symlink breakout.

  • sensitive.mjs: Blocks .env*, SSH keys, and certificates; automatically redacts OpenAI API keys, GitHub tokens, AWS credentials, and Bearer tokens.

  • Egress Sanitization: Ensures every character crossing the browser boundary—including Git diffs and execution test output—passes through an ultimate redaction filter.

• 🔄 Closed-Loop Independent Verification:

  • Automatically captures budgeted git diff outputs and persistent test execution records.

  • Anti-flattery enforcement: Without empirical execution evidence, ChatGPT will actively refuse blind approval and issue an actionable [CHANGES REQUESTED] report.

• 🔒 Single-Flight Concurrency Mutex:

  • Built-in AsyncMutex serializes concurrent MCP requests targeting the same Chrome tab, preventing prompt interleaving and signal crossing.

• 🎯 Five Specialized Operational Modes:

  1. ask: Everyday technical questions and conversational reasoning.

  2. plan: Structured planning with RATIONALE, ACTIONS, RISKS, and CRITERIA.

  3. review: Closed-loop verification of real Git diffs and execution evidence.

  4. derive: Pure mathematical and algorithmic derivation (code files isolated to eliminate noise).

  5. diagnose: Root-cause failure analysis with minimal reproduction steps.

🏗️ Four-Layer Architecture

┌─────────────────────────────────────────────────────────┐
│                 Antigravity 2.0 / Gemini                │
└───────────────────────────┬─────────────────────────────┘
                            │ JSON-RPC 2.0 stdio / CLI
┌───────────────────────────▼─────────────────────────────┐
│ 1. Access Layer (scripts/)                               │
│    - ask_chatgpt.mjs: Thin, expressive CLI entrypoint   │
│    - mcp_server.mjs: Antigravity 2.0 Native MCP Server  │
├─────────────────────────────────────────────────────────┤
│ 2. Orchestration & Protocol (src/brain/)                │
│    - orchestrator.mjs: Proxy-Pull Context Broker        │
│    - prompts.mjs: [ANTIGRAVITY-BRIDGE/1] 5 Mode Envelopes│
├─────────────────────────────────────────────────────────┤
│ 3. Security & Evidence Layer (src/security/, workspace/)│
│    - path_guard.mjs: Canonical Realpath Traversal Guard │
│    - sensitive.mjs: Secret Redaction & .env Blocking    │
│    - ignore.mjs: .brainignore Rule Parser               │
│    - context_provider.mjs: Safe Read-Only Slicing       │
│    - git_helper.mjs: Read-Only Budgeted Git Diff        │
│    - recorder.mjs: Execution Evidence Recorder          │
├─────────────────────────────────────────────────────────┤
│ 4. Transport Layer (src/transport/)                     │
│    - cdp_transport.mjs: Native CDP Driver (Port 9222)   │
│      * AsyncMutex single-flight serialization           │
│      * Base64 fast ProseMirror text injection           │
│      * Double-hash multi-sampling + MutationObserver    │
│      * Event-driven WebSocket lifecycle & cleanup       │
└───────────────────────────┬─────────────────────────────┘
                            │ Chrome DevTools Protocol
┌───────────────────────────▼─────────────────────────────┐
│ Dedicated Chrome Profile (127.0.0.1:9222)               │
│ └── ChatGPT Web UI (https://chatgpt.com/)               │
└─────────────────────────────────────────────────────────┘

🛠️ Configuring MCP in Antigravity 2.0

Google Antigravity 2.0 natively integrates external tools via the standard Model Context Protocol (MCP).

Step 1: Prerequisites

• Node.js: >= 22.0.0 (uses global native WebSocket and fetch—no npm install needed).

• Google Chrome: Standard desktop installation.

Step 2: Start the Dedicated Chrome Session

To keep your daily browser data separate, launch Chrome with an isolated user profile (chrome-profile) and local CDP remote debugging on port 9222:

Windows PowerShell:

& "C:\Program Files\Google\Chrome\Application\chrome.exe" `
  --remote-debugging-port=9222 `
  --remote-debugging-address=127.0.0.1 `
  --remote-allow-origins=* `
  --user-data-dir="D:\ChatGPT-Brain-Bridge\chrome-profile" `
  https://chatgpt.com

Tip: On first launch, log in to your ChatGPT account once. Your session will be permanently preserved in chrome-profile, requiring no future logins.

Step 3: Configure Antigravity 2.0 Global MCP Config

Edit your global MCP configuration file:

• Windows: C:\Users\<YourUsername>\.gemini\config\mcp_config.json

• Linux/macOS: ~/.gemini/config/mcp_config.json

Add antigravity-with-chatgpt under mcpServers:

{
  "mcpServers": {
    "antigravity-with-chatgpt": {
      "command": "node",
      "args": [
        "D:\\ChatGPT-Brain-Bridge\\gemini-skill\\antigravity-with-chatgpt\\scripts\\mcp_server.mjs"
      ]
    }
  }
}

Note: Update the path to match where you cloned the repository. Remember to use double backslashes \\ on Windows.

Step 4: Verification in Antigravity 2.0

Restart Antigravity 2.0. The agent will discover the tools automatically. You can instruct the agent naturally:

"Please ask ChatGPT in plan mode to design a high-throughput message queue architecture for this service."

Antigravity will dispatch the MCP tool call:

{
  "name": "ask_chatgpt",
  "arguments": {
    "mode": "plan",
    "prompt": "Design a high-throughput message queue architecture"
  }
}

💻 CLI Usage

You can also interact with the bridge directly from your terminal:

# 1. Ask a question (reuses active web session)
node scripts/ask_chatgpt.mjs "Explain the differences between Linux epoll LT and ET modes"

# 2. Start a fresh conversation (--new)
node scripts/ask_chatgpt.mjs --new "Evaluate Raft vs Paxos for distributed consensus"

# 3. Architectural Planning Mode (--plan)
node scripts/ask_chatgpt.mjs --plan "Refactor the authentication module to stateless JWT"

# 4. Closed-Loop Review Mode (--review, auto-injects real Git Diff & test evidence)
node scripts/ask_chatgpt.mjs --review "Review recent changes for security regressions"

# 5. Algorithmic Derivation Mode (--derive, noise-free math reasoning)
node scripts/ask_chatgpt.mjs --derive "Derive the variance formula for consistent hashing virtual nodes"

# 6. Root-Cause Diagnostics (--diagnose)
node scripts/ask_chatgpt.mjs --diagnose "Analyze ECONNRESET errors during test execution"

# 7. Attach Local Source Files (--attach, with automatic path guard & secret redaction)
node scripts/ask_chatgpt.mjs --attach src/auth.ts src/server.ts "Analyze thread-safety"

# 8. Run Environment Doctor (--doctor)
node scripts/ask_chatgpt.mjs --doctor

🧰 MCP Tool Reference

🧪 Testing & Verification

Run the comprehensive test suite:

# Run baseline installation & component verification
node scripts/verify_install.mjs

# Run adversarial security & concurrency tests
node tests/security_adversarial.test.mjs

Results:

• ✅ 48 / 48 Tests Passed (100%)

  • Path traversal, NUL injection & case-sensitivity: 5/5

  • Multi-line PEM, spaced secret assignment & Bearer token redaction: 5/5

  • Egress sanitization leakage prevention: 2/2

  • Orchestrator defensive contract validation: 2/2

  • Core component integrity & CDP live probing: 34/34

📄 License

Open-sourced under the MIT License. Contributions, feedback, and pull requests are warmly welcome!