Does Browser Use (MCP mode) work with Claude Desktop?
Not in our test. Browser Use (MCP mode) installed in Claude Desktop but failed our agent task on Oct 14, 2026. Details of what went wrong are below.
Result
ResultFailed
Date testedOct 14, 2026
Plan tested onNot recorded for this run
App versionClaude Desktop 2.37337.1 (Linux beta)
Steps
- install OK
- connected OK
- tools visible OK
- agent task: click reveal Failednavigate + interact + read back. The code is computed by page JavaScript only after the click.Why it failed: no tool result containing ['TP-4242']; answer missing ['TP-4242']
Test note: server wrapped in harness/scripts/mcp_tee.py to record JSON-RPC; prompts typed via xdotool; tool Allow pop-ups approved by the operator for test servers only; model = app default
From our own harness run, not the maker’s claims. How we test.
What happened in this run
Install config we used
{
"command": "uvx",
"args": [
"--from",
"browser-use[cli]",
"browser-use",
"--mcp"
]
}Connection check
initialize seen in tee log after app restartTools check
tools/list answered (tee log)
Agent task: click reveal
- 12 calls reached the server
- Agent’s answer (excerpt):
I couldn't get the code: the browser-use tools aren't working (Chrome exited before its debugging port was available; content extraction needs OPENAI_API_KEY).
Raw details from our harness log for this run. Secrets redacted.
Set it up in Claude Desktop
Copy-paste config for every app is on the Browser Use (MCP mode) page. Not connecting? Claude Desktop MCP troubleshooting.
Other apps we tested it in: Claude Code (failed) · Cursor (failed) · ChatGPT (Codex) (failed).