An Agent Browser That Runs Real Chrome

An agent browser is a real web browser that an AI agent drives toward a goal, instead of a person clicking through it or a script replaying selectors. The terms agent browser and browser agent are used interchangeably. In practice it has two layers: the agent that decides what to do, and the browser infrastructure it runs on. TestMu AI Browser Cloud is that infrastructure, providing real, full-featured Chrome sessions on demand for whichever agent is driving.

The first two are the same idea said two ways. It helps to separate three layers. A consumer agentic browser is an end-user application that browses on your behalf. An agent browser or browser agent is the programmable layer an engineer or an AI coding agent drives. Underneath both sits the browser infrastructure that actually runs the sessions, which is Browser Cloud. This page is about that infrastructure layer and the agents that connect to it.

Browser Cloud is agent-agnostic by design. Claude, Cursor, Gemini, OpenAI Computer Use, and custom in-house agents all work, because the integration surface is the Session SDK, an installable agent skill, and an MCP server rather than a per-vendor connector. If your agent can create a session, drive it, and release it, it can use Browser Cloud.

Through the built-in tunnel. Much of the work agents need to do happens before anything is public: a dev server on localhost:3000, a pre-production deployment, an internal dashboard, or a service inside a VPN. The tunnel ships inside Browser Cloud, so a cloud session reaches those targets without you exposing them to the internet or wiring up a third-party tunnelling tool. Any non-public target requires it.

Every session automatically records video, console logs, network request and response logs, and a step-by-step command replay, with no instrumentation on your side. Because agents act on what they perceive, the video and command replay answer the question headless logs cannot: why the agent decided to do that. You watch the rendered state it saw, then fix the prompt, the wait condition, or the flow.

Yes. Sessions are provisioned on demand and run independently, so many real browsers operate at once rather than queueing through one machine or a fixed local pool. Agent workloads are bursty, so throughput scales with the work instead of with hardware you pre-purchased. Per-plan session and concurrency limits are published on the Browser Cloud page.