Scalebrowser

Scalebrowser is self-hosted browser infrastructure for AI agents, running persistent browser profiles on your own hardware rather than a vendor's cloud. It ships as two components — a daemon binary that manages profiles and serves every interface, and a signed Chromium fork that profiles actually run in — so credentials, cookies, and fingerprint data never leave your machine. The product targets automation engineers and AI agent builders who need agents to stay logged in across runs, handle two-factor authentication, solve captchas, and operate multiple isolated identities simultaneously.

What It Is

Scalebrowser is a desktop-installed daemon that creates and manages browser profiles for AI agents. Each profile carries a persistent identity derived deterministically from a seed value, meaning the same profile presents as the same device on every restart. The daemon enforces coherence checks before any profile launches — if the GPU is a software rasterizer or fingerprint values contradict each other, the launch is refused. Agents connect to running profiles over MCP, direct CDP, a REST API, or an AdsPower-compatible adapter, depending on whether an AI model or custom code is driving the session.

How the Identity Layer Works

The core value proposition is that each profile is a complete, isolated identity:

• Persistent fingerprint: GPU, screen, and hardware values come from the real machine; a profile that contradicts itself refuses to start.
• Stored logins: Passwords are encrypted on disk and typed into forms; the agent never sees the plaintext value.
• Per-profile inbox: Every profile has its own email address for confirmation codes, with no external email account required.
• Two-factor codes: TOTP secrets are computed inside the profile; SMS codes arrive through the same inbox path.
• Passkeys and session cookies: Survive crashes and window closures; permission prompts are answered automatically.

The site states that zero credentials or cookies leave the machine during a run.

Agent Integration and MCP Support

Scalebrowser exposes an MCP server that Claude, Cursor, ChatGPT, Codex, Copilot, Lovable, and other MCP-compatible agents can point at over HTTP or stdio. Agents receive a handle and a small tool set rather than a raw browser connection; every click and keystroke passes through a human input layer. The site also documents a Python SDK, a Node SDK, and direct CDP access for teams that want to keep existing automation code and swap only the browser underneath. An AdsPower adapter lets scripts written against that local API continue working without a rewrite.

The site claims the token budget approach — where the agent reads the page once and then only receives diffs of what changed — costs significantly less in model tokens compared to alternatives over long runs, citing a specific figure for a hundred-step session.

Captcha and Detection Handling

Scalebrowser's site describes solving 18 captcha types on live widgets with real keys, without a third-party solving service, including press-and-hold, image grids, sliders, rotation, and audio challenges. The site also publishes an audit bench result of 8/8 detector sites passed. Blog posts dated September 2026 cover bot detection layers, browser fingerprinting check groups, and a comparison of six browser MCP servers measured on live pages for tokens per step.

Deployment Model and Requirements

Scalebrowser is self-hosted and Windows-only at launch, with no hosted or cloud mode. The daemon and engine install to a single directory; backing up that directory preserves all profiles, sessions, and the database. Three hard requirements apply in practice: your own hardware, a real GPU (software rasterizers are blocked at the pre-launch coherence check), and a residential or mobile proxy (the daemon does not force one, but running without a proxy exposes the host machine's IP before any in-browser work can help). Plans differ by how many browsers run concurrently and across how many machines; all plans include unlimited profiles and the full feature set. The Fleet plan extends at a published per-unit price with no sales call required.