Connects AI agents directly to a real, authenticated Google Chrome browser, enabling navigation, clicking, typing, screenshots, DOM snapshots, visual grounding, and other browser automation tasks using existing cookies and sessions.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Give your AI coding agents direct, ultra-fast access to your real logged-in Chrome browser.
Chrome Browser Bridge connects Antigravity CLI / IDE, OpenAI Codex, and Claude Code directly to your active Google Chrome session over a high-performance local MCP server. No separate headless browser instances, no losing login sessions, and zero unnecessary turn round-trips.
Preserve Real Logged-In Sessions: Automate tasks on authenticated sites (GitHub, AWS, WhatsApp, internal dashboards) seamlessly using your everyday Chrome profile.
Sub-Second Execution Protocol: Eliminates multi-turn LLM latency. Execute multi-step interactions, form fills, assertions, and extractions in one single compound MCP call (chrome_workflow / chrome_act).
Zero-Round-Trip Page APIs: Dispatch concurrent authenticated HTTP requests directly from page context via chrome_page_fetch_many and chrome_api_sequence.
Smart Adaptive Memory: Automatic failure recovery, in-tab human authentication checkpoints (chrome_wait_for_user), and proactive DOM observers.
Battle-Tested & Lightweight: Zero heavy external browser dependencies (like Puppeteer/Playwright); only uses native Chrome Extension APIs + lightweight WebSocket transport.
Related MCP server: OpenChrome
Node.js 22+
Google Chrome
git clone https://github.com/rahulbomnalli/chrome-browser-bridge.git
cd chrome-browser-bridge
npm installnpm startThe local bridge server runs on http://127.0.0.1:8765.
Open Chrome and navigate to chrome://extensions/.
Toggle on Developer mode in the top-right corner.
Click Load unpacked and select the extension/ folder inside this repository.
Verify the setup in a new terminal:
node cli.js doctorRun the configuration generator:
node cli.js config antigravityMerge the generated output into your global config (~/.gemini/config/mcp_config.json) or workspace configuration (.agents/mcp_config.json). Generated Antigravity, Claude, and Codex configs use the task-focused fast profile: 33 tools covering universal browser controls, workflows, extraction/API work, visual capture, downloads, page diagnostics, and change monitoring, with compact successful results. Set BRIDGE_TOOL_PROFILE=compact or BRIDGE_TOOL_PROFILE=full to expose the broader catalogs.
Add to your project .mcp.json or run:
claude mcp add --env BRIDGE_TOOL_PROFILE=fast --transport stdio chrome-browser-bridge -- node "/absolute/path/to/chrome-browser-bridge/mcp-server.js"Run:
codex mcp add chrome-browser-bridge --env BRIDGE_TOOL_PROFILE=fast -- node "/absolute/path/to/chrome-browser-bridge/mcp-server.js"Actions that expose timeoutMs or timeoutSec accept per-call deadline overrides. Set BRIDGE_DEFAULT_TIMEOUT_MS to change the fallback deadline; human checkpoints and page-watch actions retain their dedicated defaults.
The server defaults to the 33-tool fast profile, and generated client configs set it explicitly. Alongside compound actions, it now exposes universal tab, navigation, snapshot, click, type, scroll, select, wait, and overlay controls. It also includes parallel URL research, batched extraction/API requests, screenshots, clean page reading, downloads, network/console diagnostics, page-change monitoring, idempotent browser.execute_plan, cancellation/status checkpoints, and bounded execution metrics. Prefer chrome_act or chrome_workflow to batch known multi-step work. Specialized CDP and macro tools remain available through BRIDGE_TOOL_PROFILE=compact or BRIDGE_TOOL_PROFILE=full.
MCP guidance requires clients to ground claims in returned tool data, inspect unknown page state instead of guessing, and report task success only when the result and final verification support it. Legacy tools return structured browser results rather than unconditional success summaries.
Tool | Purpose | Performance Benefit |
| Multi-stage tasks spanning navigation and conditional logic | Executes full plan locally in browser; zero model round-trips |
| Execute a validated plan with idempotency and persisted checkpoints | Prevents duplicate submissions and serializes same-tab plans |
| Cancel at a stage boundary or inspect saved task state | Safe recovery without blind replay |
| Read MCP and bridge latency/outcome traces | Separates bridge cost from unavailable model/client time |
| Batch same-page interactions (click, type, assert, extract) | Single-injection script with native setters and microtask reactivity |
| Read independent known URLs concurrently | One call for multi-site page research |
| Switch to/reuse already-open tabs or navigate active tab | Eliminates duplicate tab churn and preserves SPA state |
| Control tabs and navigate directly | Generic browser controls for simple navigation tasks |
| Inspect and perform one grounded interaction | Direct controls when a compound action is unnecessary |
| Scroll, select a dropdown option, or wait for page state | Covers common single-step interactions |
| Dismiss common banners and dialogs | Unblocks page interaction without site-specific tooling |
| Execute 1–20 API requests in active page session | Uses existing browser session cookies/headers concurrently |
| Execute dependent API requests in one page injection | Reuses prior response values without model round-trips |
| Inspect requests and resources already observed on the page | Finds known page APIs without guessing endpoints |
| Extract multiple CSS selectors in a single call | One round-trip extraction instead of repetitive DOM queries |
| Read clean headings and text from the active page | Useful for articles and pages without selector-specific extraction |
| Capture the visible browser tab | Provides visual evidence for layout-dependent tasks |
| Read recent network and console events together | One call for routine page diagnostics |
| Download a known URL or data URI through Chrome | Handles downloads without a separate browser workflow |
| Wait for a selected page region to change | Avoids polling snapshots for asynchronous page updates |
| Auto-detect WhatsApp tab, resolve recipient, and send message | Handles App Lock and verification in one call |
| Prompt user for manual CAPTCHA / 2FA checkpoint | In-tab modal pauses agent safely until user completes verification |
Run the comprehensive suite of unit tests, syntax checks, and transport benchmarks:
# Run unit & integration tests (50/50 passing)
npm test
# Run syntax check across all modules
npm run check
# Run transport benchmark
npm run bench -- --mock├── extension/ # Manifest V3 Chrome Extension (background, DOM dispatch, workflow)
├── lib/ # Client config, transport layers, and tool profiling
├── test/ # Comprehensive unit and integration test suite
├── docs/ # In-depth architecture specs, audits, and guides
├── cli.js # Diagnostic CLI tool for status, doctor, tabs, and actions
├── server.js # Local HTTP & WebSocket bridge server
├── mcp-server.js # Fast Model Context Protocol (MCP) server implementation
└── package.json # Dependencies and scriptsThis project is licensed under the MIT License.
This server cannot be deployed
Let your AI agent operate any website through a real, logged-in browser session.
Connect AI agents to your own Chrome with scoped access, approvals, and verified task results.
Give AI agents a real cloud Chrome: open sessions, navigate, read pages, click, type and screenshot.
Control a real browser from AI agents: navigate, click, type, screenshot, read the page.
