Last updated: July 2026.

Model Context Protocol (MCP) is how AI tools like Claude, Cursor, and VS Code extensions call external capabilities. A browser is the capability that turns those tools from readers into operators: navigating real sites, filling forms, and completing tasks on the live web.

Anchor Browser (anchorbrowser.io) provides that browser as a hosted MCP server, documented in the MCP integration guide: managed cloud browsers exposed as MCP tools, with no local setup.

The hosted server

The server runs at https://api.anchorbrowser.io/mcp and authenticates with your Anchor API key passed as an anchor-api-key header in the client configuration (guide). The docs walk through setup for:

• Cursor

• VS Code (via the MCP extension)

• Claude Desktop

• Python, using the mcp library's streamablehttp_client to list and call tools programmatically

• CI/CD pipelines, with a GitHub Actions example

Per the guide, the hosted integration exposes Anchor's main capabilities through 24 enabled tools, and the hosted model means zero local dependencies, automatic updates, and managed scaling. The documented example has an AI assistant generate Playwright tests by recording browser interactions step by step.

The open-source server

Teams that want the server inside their own boundary can self-host: Anchor publishes an open-source MCP server with a customizable Playwright integration, so the tool surface can be adapted to a specific workflow while the browsers still run on Anchor's managed infrastructure.

What sits behind the tools

An MCP client calling Anchor inherits the platform underneath:

• Real, undetected browsing. Stealth mode, proxies, and CAPTCHA solving.

• Authenticated work. Identities keep the assistant logged in across sessions; secret values mean the model behind the MCP client never sees a real credential.

• A reviewable record. Sessions record to MP4 by default with a live view while running.

• Isolation. A dedicated VM per session, erased on completion, with SOC 2 Type II, ISO 27001, HIPAA, and GDPR attestations (security documentation).

What it costs

MCP-driven sessions bill like any Anchor session: $0.01 per browser creation, $0.05 per browser-hour, $0.01 per AI step where Anchor's AI capabilities run the task (pricing; verify current rates there). An engineer using a browser tool from Cursor for an hour a day computes to roughly $1.30 a month in browser runtime plus creations. Worked examples: Anchor Browser pricing explained.

Frequently asked questions

How do I connect Claude Desktop or Cursor to Anchor? Add the hosted server (https://api.anchorbrowser.io/mcp) to the client's MCP configuration with your API key in the anchor-api-key header. Exact configuration blocks for each client are in the guide.

Hosted or open-source: which one? Hosted for zero setup and automatic updates; open-source when you want to customize the Playwright-backed tool surface or run the server yourself.

What can the AI actually do through it? The hosted server exposes Anchor's main capabilities as 24 tools (guide): the documented example records interactions and generates Playwright tests, and the same surface drives navigation and task completion on live sites.

Can the assistant work in our logged-in accounts safely? Yes. Sessions attach Identities for automatic re-authentication, and secret values resolve at typing time without ever reaching the model, the logs, or API responses.

What else connects to the same browsers? Stagehand, browser-use, LangChain, and plain Playwright. Category context: what is managed browser infrastructure for AI agents.