Isolated browsers with a persistent identity for AI agents, on your own Windows machine.
net.scalebrowser/mcp MCP Server
The net.scalebrowser/mcp Model Context Protocol (MCP) server provides isolated browser instances with a persistent identity for AI agents running on your own Windows machine. Its description emphasizes local execution and continuity of identity across agent interactions.
🛠️ Key Features
Isolated browsers
Persistent identity for AI agents
Runs on your own Windows machine
🚀 Use Cases
Running AI agents in isolated browser environments
Maintaining consistent agent identity across browser sessions on Windows
⚡ Developer Benefits
Browser isolation for agent workflows
Persistent identity to support continuity during agent operations
Your agent got blocked, stopped at a login, or gave up on a captcha? Scalebrowser gives each AI agent its own isolated browser with a persistent identity, running on your own machine instead of a vendor's cloud. It stays signed in, handles the captcha, types like a human, and you can take over whenever you want.
This repository connects that browser to your agent over MCP: copy-paste configurations for every client the documentation covers, a small stdio bridge, and runnable SDK examples.
Before you start
Windows 10 or 11, 64-bit, with a real GPU. Scalebrowser is a desktop app that runs on your machine. There is no hosted mode.
A Scalebrowser account with an active plan. The way in is a 7-day trial with one browser. A payment method is required up front. Plans and prices
Free, no account needed:scalebrowser.net/check measures how your current browser looks from the outside, and scalebrowser.net/ai-agent-check measures what your automation tool does with it. Point your agent at the second one and see what a website can tell.
Try it in five minutes
Sign in at scalebrowser.net, open Download, and run the installer. It installs for the current user and asks for no administrator rights. SmartScreen warns on first run: choose More info, then Run anyway.
Start the app and press Sign in. The app then downloads and verifies the browser engine, which has to finish before any profile can start.
Copy your API token from Settings, General, Connect an agent.
Connect your agent. For Claude Code it is one command:
Open news.ycombinator.com in a Scalebrowser profile, read the front page, and tell me the three stories with the most comments.
The agent takes a lease, which reserves a profile and starts its browser, opens the page, reads a text map of it and answers. Then ask it to release the profile.
Connect your client
Each file below is the configuration from the official documentation, ready to copy. Replace <token> with your API token and <you> with your Windows user name. Where each file goes, and the two spellings that fail without an error, is in clients/README.md.
Claude on the web, Grok, ChatGPT and other cloud chat services
No file: they reach your machine through remote access
What your agent gets
The agent works from a text map of the page and addresses elements by reference. It looks at the page once, then only at what moved. Over 75 steps on the same page, that spends 11x less on model tokens than Playwright MCP.
Signing in:credential_fill signs in with a login stored in the profile, and credential_new signs up with a password Scalebrowser draws and stores. The agent never sees the value.
Confirmation codes:read_inbox reads the codes sent to the address bound to the profile, and inbox_open_link opens a link from a message.
Challenges:press_and_hold, click and four measuring tools work through verification challenges with the same human input. The measured status per challenge type says what passes and what does not.
Proof of the work:record and video turn a run into a film, and every tool call becomes a step in the run's trail.
The full list is in the tool reference. What an agent may do is decided by the app, not by the client.
The bridge
bridge/scalebrowser-mcp.mjs is a stdio MCP server with no dependencies. With the Scalebrowser app running on the same Windows machine, it relays every message to the app's own MCP endpoint, using the address and token the app keeps in %LOCALAPPDATA%\Scalebrowser. Without the app, it answers with the tool list and install instructions. That second mode is what MCP directories see when they build this repository.
Clients that speak HTTP do not need it: use the configurations above. For a client that only starts servers over stdio, run it from npm:
bash
npx -y @scalebrowser/mcp
or from a clone of this repository:
bash
node bridge/scalebrowser-mcp.mjs
It needs Node 22 or newer. --data-dir points it at another data directory, and SCALEBROWSER_TOKEN overrides the token file.
As a Claude desktop extension
Each release carries the bridge as an MCP Bundle, scalebrowser-<version>.mcpb. Download it on the Windows machine the app runs on and open it with the Claude desktop app, or choose Settings, Extensions, Install Extension there. It needs no configuration: it finds the app and its token by itself.
The bridge is also listed in the MCP Registry as net.scalebrowser/mcp.
As a Claude Code plugin
This repository is also a Claude Code plugin. In Claude Code on the Windows machine the app runs on:
The plugin starts the bridge, so it needs Node 22 or newer and no token in any file. It also adds a skill that teaches the agent the lease, read, act, release loop, to look before it repeats a failed action, and to sign in with stored logins instead of asking for a password.
SDK examples
The same small task in both SDKs: open Hacker News in a profile and print the three top stories. Run them on the Windows machine the app runs on, with your token in SCALEBROWSER_TOKEN.
bash
cd examples/node
npm ci
node top-stories.mjs
bash
cd examples/python
pip install --require-hashes -r requirements.txt
python top_stories.py
The bridge collects nothing, stores nothing and sends nothing to us or to anyone else.
What it reads: the address and API token the Scalebrowser app keeps in its data directory on your machine, %LOCALAPPDATA%\Scalebrowser unless you pass --data-dir.
Where messages go: only to the Scalebrowser app on the same machine. It refuses any address that is not this machine's own loopback address, so the token and your agent's messages never leave the computer through the bridge.
What it keeps: nothing. It holds the session with the app in memory and ends it when your client disconnects. It writes no files and no logs besides the status lines it prints to standard error.
Third parties: none. The bridge has no dependencies, no telemetry and no network access of its own beyond the loopback connection to the app.
