02 / 06
.
MCP agent integration with live browser control
- Category
- AI Agents · Automation
- Role
- Agent integration & automation
- $ claude --mcp browser
- → mcp server connected · live browser session
- ✓ script_01 typed i/o contract pass
- ✓ script_02 typed i/o contract pass
- ✓ script_03 typed i/o contract pass
- ✗ 3/3 failed cold env · window size drift
- → tracing shared downstream dependency…
- ✓ fixed · re-validated from cold start 3/3
- $
By the numbers
- 3
- typed automation scripts
- MCP
- live browser bridge
- 3/3
- re-validated from cold start
Overview
I connected Claude Code to a third-party MCP server so the agent could control a live browser, a capability it did not have on its own. Over that connection I built three automation scripts, each with a typed input and output contract, so every run could be checked against what it promised to return.
The regression
All three scripts, already validated and deployed, suddenly failed together. Reading the run traces showed why: the platform’s cold environment opened at an inconsistent window size, which silently broke a dependency the three scripts shared downstream. I fixed it, then re-validated from a fresh cold start rather than a warm session, so the fix was proven where the bug actually lived.
Highlights
- 01Bridged Claude Code to a third-party MCP server for live browser control.
- 02Three automation scripts, each with a typed input/output contract.
- 03Root-caused a shared regression from run traces: inconsistent window size in the cold environment.
- 04Re-validated from a fresh cold start, not a warm session.
- 05Wrote structured platform feedback: reproduction steps, root cause, severity and a suggested fix.