Is anyone else losing their mind testing MCP servers across different clients?
Hey all,
I have been building MCP servers for a few months now and the cross-client testing situation is killing me. I wanted to see if this is just a skill issue before I go build something.
Last week I had an MCP that worked perfectly in Claude Desktop, exposed all 12 tools in Gemini CLI, but in Copilot CLI it showed zero tools. No error, no warning, no log entry. Just silent failure. It took me about 4 hours of adding print statements everywhere to figure out it was choking on a specific JSON schema field that Copilot's client parses more strictly than the others.
This keeps happening. I found out that different LLMS have different tolerances for schema quirks, different auth handshakes, different timeout behaviors.
So I have been thinking about building something that runs your MCP server against a bunch of real agent clients (Claude Desktop, OpenAI Agents SDK, Gemini, etc.) and tells you where it breaks and why. It may run as a CLI or a GitHub Action on every PR.
Before I start developing this, I would love a sanity check:
- Is cross-agent MCP compatibility an issue everyone is facing, or am I doing something wrong? If so, then what am I doing wrong?
- What breaks most often in your experience? Connection, tool discovery, execution, auth, or client bugs?
- How long does debugging usually take when an MCP works in one client but not another?
Thanks for any feedback!