Controlled test rig for MCP client timeout and streaming behaviour. Spins up a configurable local MCP server (slow tools, progress notifications, SSE keep-alives, header logging) and runs headless Claude sessions against it to verify how the client handles long-running tool calls.
Or install with pip: pip install mcp-client-testbench
README Excerpt
**The question this repo answers:** when an MCP tool call takes a long time, what does the Claude client actually do, and does anything the server sends (progress notifications, keep-alives) stop it from timing out? **Why it exists:** Claude Code and Claude Cowork abort MCP tool calls that exceed a timeout (observed: 60 seconds for claude.ai connectors). Server