browser-gateway

browser-gateway

The open router for cloud browsers.
One endpoint that routes across every browser provider you use: automatic failover, persistent profiles, session replay, REST API, MCP server, dashboard.
Works unchanged with Puppeteer, Playwright, Stagehand, browser-use, and any MCP client.

     

Website  ·  Quick start  ·  MCP  ·  Profiles  ·  Replays  ·  REST API  ·  Dashboard  ·  Discord

---

Overview

One endpoint. Multiple providers. Automatic failover when one is saturated or goes down.

Your app connects to ws://gateway:9500/v1/connect. The gateway picks the best available provider based on health, capacity, and your routing strategy. Providers can be cloud CDP services, Docker containers, or local Chrome instances.

---

Dashboard

A web dashboard ships with every install. Open http://localhost:9500/web after starting the gateway.

Overview. Active sessions, queue depth, provider health, connection endpoint, and a copy-paste quickstart for Puppeteer, Playwright, Stagehand, browser-use, and raw CDP.

REST API. Run screenshot, content extraction, and structured scraping endpoints from a form-driven UI, with profile selection and parameter reference inline.

Playground. Drive any provider live from the browser. Pick a provider and profile, type into the canvas as if it were a local browser, and watch the remote session in real time.

---

Features

Routing & reliability

• Automatic failover - the next provider takes over the instant one fails, no client changes
• Five load-balancing strategies - priority chain, round-robin, least-connections, latency-optimized, weighted
• Per-provider concurrency limits - the gateway enforces maxConcurrent on every backend
• Request queue - connections wait when every provider is saturated instead of failing immediately
• Cooldown - failing providers are skipped and recover automatically after a TTL
• Health checks - periodic connectivity probes mark providers unhealthy before clients hit them
• Graceful shutdown - active sessions drain cleanly on SIGTERM and SIGINT
• Session reconnect - a dropped client reconnects with ?sessionId=<X-Session-Id>; on a provider that issues a resume token (browserserve with its resume window on) it gets the same browser back with its pages, cookies and storage, otherwise a new browser on the same provider
• Webhooks - fire on provider down, recover, and queue-overflow events

REST API

• Screenshot - POST /v1/screenshot returns any URL as PNG or JPEG, full-page or scoped to a selector
• Content extraction - POST /v1/content returns markdown, plain text, HTML, or a cleaned article
• Scrape - POST /v1/scrape extracts structured data via CSS selectors or full-page formats
• Pooled sessions - browser connections are reused across requests, like a database pool
• Automatic retry - failed requests retry against a fresh page

Profiles — persistent browser state

• Survive across sessions - cookies and localStorage are captured as you leave each page and on disconnect, and replayed on the next connect with the same id (IndexedDB too on browserserve)
• Saved on any provider - every provider can load and save any profile; pin a provider to one profile only to route that profile's sessions there
• Profiles never mix - a browser that still holds a different profile is refused. Anything else already in a reused browser (from sessions without a profile) is saved with the profile
• One-line opt-in - add ?profile=acme to the WebSocket URL, the rest is automatic
• Encrypted at rest - AES-256-GCM with envelope encryption, anti-swap binding, and a scrypt-derived KEK
• Provider-agnostic - state is captured at the CDP level, so it replays against any provider
• Per-profile locking - concurrent connects to the same id return HTTP 409 to prevent corruption
• Export and import - encrypted .bgp blobs are portable between gateway installs
• One-click enable - the dashboard wizard generates a strong key in your browser and writes it to config

See the Profiles docs for the full guide, security model, REST endpoints, and limitations.

Session replay — see what the agent saw

• Frame-accurate visual record of every routed session, captured via CDP Page.startScreencast
• Zero injection - no script runs inside the customer page, capture is fully out-of-band
• Provider-agnostic - works with any backend that supports page screencast
• Dashboard player - scrub through the recorded frames, switch between captured browser targets
• Retention controls - configurable horizon, per-session byte cap, daily cleanup

See the Replays docs for the storage layout, REST endpoints, and tuning knobs.

MCP server for AI agents

• One-call tools - fetch_page reads any url as markdown and screenshot_page captures it, each in a single call that opens and closes its own browser
• Session tools - browser_session, browser_navigate, browser_snapshot, browser_act, browser_extract, browser_screenshot, browser_wait, browser_tabs, browser_evaluate, browser_observe for logins, forms and multi-step work
• Fallback for coding agents - a Claude Code plugin sends Claude to fetch_page when its built-in web fetch is blocked, empty or needs JavaScript (/plugin marketplace add browser-gateway/browser-gateway); setup for Codex, Cursor and Gemini CLI in plugins/browser-gateway
• Zero config - auto-detects Chrome and launches it on first tool use
• Concurrent sessions - every agent gets its own browser, no shared state
• Raw CDP - no Playwright or Puppeteer dependency
• Compatible - Claude Code, Cursor, and any MCP-compatible client

Computer use for Claude and Gemini

• Any browser for the model's own browser tool - runs Claude's browser and computer use tools and Gemini computer use on any provider behind the gateway
• About ten lines - runClaudeLoop / runGeminiLoop take your own model call and handle actions, screenshots and results
• Safe defaults - http and https only, private networks blocked, optional site allowlist, Gemini's confirm step for risky actions
• No extra dependencies - browser-gateway/agent-tools/computer-use, tested on Node.js, Bun, Deno and the Cloudflare Workers runtime

Management

• Dashboard - manage providers, watch sessions, and edit config from the browser
• Provider CRUD - add, edit, delete, and test providers from the dashboard or API
• Config editor - edit gateway.yml in-browser with syntax highlighting and validation
• Auth - token-based, with a secure HttpOnly cookie for the dashboard
• Protocol-agnostic - works with Playwright, Puppeteer, and any WebSocket protocol

---

Quick Start

As a WebSocket Proxy (for applications)

npm install -g browser-gateway

Create gateway.yml:

version: 1

providers:
  primary:
    url: wss://provider.example.com?token=${PROVIDER_TOKEN}
    limits:
      maxConcurrent: 5
    priority: 1

  fallback:
    url: ws://my-playwright-server:4000
    limits:
      maxConcurrent: 10
    priority: 2

browser-gateway serve

Connect from your app:

// For CDP providers
const browser = await chromium.connectOverCDP('ws://localhost:9500/v1/connect');

// For Playwright run-server providers
const browser = await chromium.connect('ws://localhost:9500/v1/connect');

Or use the REST API — no WebSocket management needed:

# Screenshot
curl -X POST http://localhost:9500/v1/screenshot \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com"}' --output screenshot.png

# Extract content as markdown
curl -X POST http://localhost:9500/v1/content \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com", "formats": ["markdown"]}'

These endpoints accept http and https URLs on public hosts only. Loopback, link-local and private-range addresses are refused, and hostnames are resolved and re-checked. To reach an internal host, list it under rest.allowedPrivateHosts in gateway.yml.

Dashboard at http://localhost:9500/web.

As an MCP Server (for AI agents)

Add to your Claude Code or Cursor config:

{
  "mcpServers": {
    "browser-gateway": {
      "command": "npx",
      "args": ["browser-gateway", "mcp"]
    }
  }
}

No config files needed. The agent gets navigate, snapshot, screenshot, click, type, and evaluate tools through the gateway's routing layer.

See the MCP docs for all options.

---

Authentication

Set BG_TOKEN to require a token (or put it in a .env file):

BG_TOKEN=my-secret-token browser-gateway serve

• WebSocket clients pass the token as ?token= query param
• API clients use Authorization: Bearer <token> header
• Dashboard shows a login form, sets a secure HttpOnly cookie
• Health endpoint (/health) is always public

Without BG_TOKEN every /v1/* route is unauthenticated, so the gateway binds to 127.0.0.1 instead of all interfaces. Set a token (or set HOST explicitly) to serve other machines. BG_ALLOWED_HOSTS allows extra Host values on /mcp when no token is set.

---

CLI

# Proxy server
browser-gateway serve                    # Start the gateway + dashboard
browser-gateway serve --port 8080        # Custom port
browser-gateway serve --config path.yml  # Custom config

# MCP server for AI agents
browser-gateway mcp                      # Auto-detect Chrome, zero config
browser-gateway mcp --headless           # Headless mode (for CI/Docker)
browser-gateway mcp --cdp-endpoint ws:// # Connect to existing browser
browser-gateway mcp --config gateway.yml # Multi-provider with failover

# Utilities
browser-gateway check                    # Test provider connectivity
browser-gateway version                  # Print version
browser-gateway help                     # Show help

---

API

Endpoint	Method	Description
/v1/connect	WebSocket	Connect to a browser (the core feature)
/v1/screenshot	POST	Take a screenshot of any URL (docs)
/v1/content	POST	Extract page content as markdown, text, or HTML (docs)
/v1/scrape	POST	Extract data via CSS selectors or full-page formats (docs)
/v1/status	GET	Gateway health + provider status + pool status
/v1/sessions	GET	Active sessions
/v1/providers	GET/POST	List or add providers
/v1/providers/:id	PUT/DELETE	Update or remove a provider
/v1/providers/:id/test	POST	Test provider connectivity
/v1/config	GET/PUT	Read or save config
/v1/config/validate	POST	Validate YAML without saving
/mcp	POST	MCP Streamable HTTP endpoint
/json/version	GET	CDP discovery (for browser-use, Playwright, Stagehand)
/health	GET	Health check

---

Docker

Recommended: Docker Compose. The bundled docker-compose.yml mounts a named volume for state and a read-only gateway.yml from the host.

# Drop your gateway.yml next to docker-compose.yml, then:
docker compose up -d

Plain docker run:

docker run -d \
  -p 9500:9500 \
  -v bg_data:/data \
  -v ./gateway.yml:/app/gateway.yml:ro \
  -e PROVIDER_TOKEN=xxx \
  ghcr.io/browser-gateway/server:latest

Persistence

Everything the gateway writes to disk lives under a single directory, BG_DATA_DIR (defaults to /data inside the image). Mount that as a named volume or a bind mount and all state survives container restarts and image upgrades. Today it contains:

• profiles/ — encrypted profile blobs (when profiles are enabled)

Future versions may add more subdirectories under the same root (cooldown state, session snapshots, captures). Mounting BG_DATA_DIR as one volume keeps every subsystem persistent without follow-up config changes.

Upgrades

State lives in the volume, code lives in the image. Pull the new image, recreate the container — no data lost:

docker compose pull
docker compose up -d

The container reads the same BG_DATA_DIR and the same gateway.yml. Profile blobs are versioned and the gateway reads older formats transparently.

Image tags

Tag	Updated on
:0.3.0 (and every subsequent version)	published manually after a release
:latest	always points at the newest version

Images are multi-arch (linux/amd64, linux/arm64), signed with Sigstore build provenance, and ship an SBOM. Verify with the GitHub CLI:

gh attestation verify oci://ghcr.io/browser-gateway/server:0.3.0 \
  --repo browser-gateway/browser-gateway

---

How It Works

Sessions without profile / recording / observability (the default) take the byte-pipe fast lane:

1. Client connects to ws://gateway:9500/v1/connect
2. Gateway selects a provider using your routing strategy
3. Gateway opens a raw TCP connection to the provider
4. HTTP upgrade forwarded, provider responds with 101 Switching Protocols
5. Bidirectional TCP pipe: client <-> gateway <-> provider
6. All WebSocket messages forwarded transparently (never parsed or modified)
7. On disconnect: session cleaned up, slot released, metrics updated
8. If all providers full: connection waits in a queue until a slot opens

Sessions with profile inject, session recording, live view, or observability (e.g. ?profile=X, ?session_record=true, /v1/live) run through a CDP-aware pipeline instead — one WebSocket per session, N plugins observing the wire. Same routing + failover, byte-perfect passthrough at rest, plugins only fire when their feature is requested. Architecture + plugin-authoring guide: docs/PIPELINE.md.

---

Self-hosted provider: browserserve

browserserve is the stack's own self-hosted browser server: one container that hands out isolated Chrome sessions over CDP. Add it like any other provider:

providers:
  browserserve:
    url: ws://your-host:9222
  cloud-provider:
    url: <websocket-url-with-auth>
    priority: 2

Because the gateway controls that runtime, a browserserve provider is auto-detected and unlocks two things no external provider gets:

• Auto capacity. You do not set maxConcurrent. browserserve measures its host (memory, thread, and CPU limits) and advertises a safe ceiling, which the gateway adopts. The dashboard shows it as (auto).
• Complete profiles. Every session is a fresh browser, and a saved profile carries the whole browser state, including IndexedDB and service workers. Other providers save cookies and site storage.

A common shape: browserserve as the primary provider on your own hardware, with a cloud provider at a lower priority for failover.

---

Works With

browser-gateway is compatible with existing browser tools. Just pass the gateway URL — it auto-resolves via /json/version.

AI Agent Frameworks:

# browser-use (Python) — HTTP URL auto-resolves
BrowserSession(cdp_url="http://localhost:9500")

// Stagehand (TypeScript)
new Stagehand({ env: "LOCAL", localBrowserLaunchOptions: { cdpUrl: "http://localhost:9500" } })

Playwright MCP (all 70 Playwright tools through gateway routing):

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest", "--cdp-endpoint", "http://localhost:9500"]
    }
  }
}

Puppeteer / Playwright:

// Playwright — HTTP or WebSocket
const browser = await chromium.connectOverCDP("http://localhost:9500");

// Puppeteer — WebSocket
const browser = await puppeteer.connect({ browserWSEndpoint: "ws://localhost:9500/v1/connect" });

---

Documentation

Full docs live at docs.browsergateway.com.

• Getting Started
• Configuration Reference
• Supported Providers
• Profiles — Persistent Browser State
• Session Replays
• Session Lifecycle
• REST API
• MCP Server for AI Agents
• Computer Use (Claude and Gemini)
• CLI Reference
• Web Dashboard
• How Failover Works
• Load Balancing Strategies
• Request Queue
• Webhooks
• Docker Deployment
• Integrations — Playwright, Puppeteer, browser-use, Stagehand, Playwright MCP

---

Contributing

Contributions welcome. See CONTRIBUTING.md for guidelines.

License

MIT - see LICENSE.

Links

• browsergateway.com
• docs.browsergateway.com
• GitHub
• npm

Contact

Questions, security reports, or partnership inquiries: [email protected].