browser-use 1.0.7 stops long MCP calls timing out at 60 seconds
The release returns partial output within 50 seconds while the browser job keeps running, so clients like Cursor get a result instead of an error.
What shipped on 20 November
browser-use 1.0.7 is out, installable with pip install 'browser-use==1.0.7'. The headline item is a browser-use MCP timeout fix: MCP calls no longer time out at 60 seconds. According to the release notes, a long browser_exec now replies within 50 seconds with the output produced so far, which means Cursor and other clients that enforce a 60 second limit get a result rather than Request timed out.
The release was published on 20 November and carries issue references #6477 for the timeout change, #6442 and #6422 for documentation updates, and #6454 for feedback prompts.
How the partial reply works
The change is about response timing, not about cutting the job short. The release says the cell keeps running after the early reply. When you want the remainder, you call browser_exec again with empty code, and that call returns the rest of the output.
Calls that finish in under 50 seconds are unchanged. They reply exactly as before, so existing short browser tasks should behave the same way on this version.
That split matters because a 50 second reply is a snapshot, not a final answer. The tool returns what it has so far and leaves the work in progress on the other side. Any workflow that consumes the reply has to be able to ask a second time rather than assume the first response is complete.
Why a 60 second limit was a problem
MCP clients commonly cap how long they wait for a tool response. Cursor is named in the notes as an example of a client with a 60 second limit. When a browser automation step runs past that cap, the client gives up and reports a timeout even though the underlying job may still be executing. From the agent's point of view, the work is lost, and the model has no output to reason about.
Returning early with partial output turns a hard failure into a resumable step. The client stays inside its own limit, and the agent gets something to read instead of an error string.
The practical pattern the release describes is a loop: call browser_exec, read whatever came back, then call it again with empty code to collect the remainder. That is only supported for the flow the notes describe. Nothing in the source says other tools or other timeout windows behave the same way.
Documentation and feedback changes
The release also lists README changes: a section titled Browser Use for enterprises, and the Cloud API key link restored in the Python quickstart. Separately, the Skills files, README, CLI and MCP now ask coding agents for integration feedback (#6454). The notes do not describe what that request looks like or where the feedback goes.
Limits and open questions
The source does not say whether the 50 second reply window can be configured, or whether other MCP tools in the package get the same treatment as browser_exec. It also does not name which client versions were tested beyond describing Cursor and similar clients with a 60 second cap, and it does not explain how a client should detect that a reply is partial rather than final.
One more detail is left implicit. The notes refer to the running job as a cell, which suggests a notebook-style execution model, but the release does not explain how that state is stored or how long it survives between calls. If the state expires, the second call with empty code would have nothing to return, and the notes do not address that case.
What this means for people building with AI
Anyone wiring browser automation into an agent through MCP has probably hit a client-side timeout and blamed the tool. This release moves the problem to the response format: instead of one call that must finish inside the client's limit, you get a partial result and a follow-up call. Builders using browser-use from Cursor or a similar client should check that their agent loop can issue that second call, because a single read of the first response will look like truncated output.
If you are assembling AI features into an app and want to see how Draftship handles model choice and the chat surface, the AI chat and models documentation covers that setup. The browser-use change itself is described in the 1.0.7 release notes.
Primary sources
- browser-use 1.0.7 · Browser Use releases, Nov 20, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.