MCP on Browserberg
A Model Context Protocol server, hosted on the control plane and open source, that exposes Browserberg sessions, the hosted task agent, workflows and the action log as tools to any MCP client — with the same safety gates on every action.
Details
| Integration status | Shipped — hosted at /v1/mcp, and as the npm package @browserberg/mcp |
|---|---|
| Licence | Apache-2.0, alongside the wire protocol and the SDKs |
| Transport | Streamable HTTP (hosted, API key in the Authorization header) and stdio (local) |
| Where sessions run | EU bare metal (OVHcloud) |
As of 2026-09-04
Connect in one line
{
"mcpServers": {
"browserberg": {
"type": "http",
"url": "https://browserberg.com/v1/mcp",
"headers": { "Authorization": "Bearer bb_..." }
}
}
}
Questions people actually ask
What can an agent do through it?
Read a page and act on it step by step, hand a whole multi-page goal to the hosted agent and get back a verified outcome, publish and schedule workflows, sign in with vault credentials by reference, and read the tamper-evident action log. Only the tools your deployment supports are listed.
Does it work with claude.ai’s remote connectors?
Not yet. The hosted endpoint authenticates with an API key in a header, which Claude Code, the Claude Agent SDK and any header-capable client can send; claude.ai’s connector UI expects OAuth, which is not offered today.
Is it open source?
Yes, Apache-2.0, alongside the wire protocol. The hosted endpoint runs the same package, handed a client with your key — it has no privileged access. The runtime that executes sessions stays closed, and we would rather state that plainly than let an open-source badge imply more than it covers.
Things worth knowing
Secrets never pass through the conversation
No tool creates or reads back a credential. Logins reference a vault entry by id; the secret is typed inside the browser, site-checked, and never returned. Webhook signing secrets are created in the dashboard for the same reason.
Open source, like the protocol
The MCP server ships with a skill that teaches an agent when to read a page itself, when to delegate, when to record a workflow, and how to read the three task outcomes — completed, terminated, failed — so a refusal is relayed rather than retried.
Other integrations
- Playwright Playwright connects to Browserberg over the Chrome DevTools Protocol.
- Puppeteer Puppeteer talks CDP natively, so connecting to Browserberg is a one-argument change from puppeteer.
- Selenium Selenium reaches Browserberg through Chrome's CDP support.
- Chrome DevTools Protocol If you would rather not use a driver library at all, a session is a plain CDP WebSocket.
- LangChain LangChain agents drive Browserberg today through Playwright's toolkit, pointed at a remote CDP endpoint.
Give your agent a browser
Five browser hours free, no card. Add the endpoint to your MCP configuration and the first tool call opens a session.