Mcpop — AI skill for Claude Code

MCPOp

Transparent observability proxy and silent failure catcher for Model Context Protocol (MCP) servers.


---

Why this exists

If you use Claude Desktop, Cursor, or autonomous agents with MCP servers, you have likely run into this:

You install a bunch of MCP tools, start a coding session, and suddenly your token budget evaporates while the agent produces garbage output.

Here is what is actually happening under the hood:

1. The Retry Death Loop: When an MCP tool fails silently (expired token, database timeout, internal 500), the agent does not just stop. It aggressively retries the exact same broken payload 5 to 10 times in a row. Because every retry re-transmits the entire conversation context window, you burn 50,000+ tokens in 30 seconds for zero result.
2. Schema Mismatches: The agent hallucinates arguments or drops a required property. The tool crashes, but because everything runs over background stdio pipes, you see zero logs.
3. Latency Black Holes: A sluggish tool (>5s) stalls the agent workflow until the client times out.
4. Painful Debug Loops: Testing a single broken tool call used to require restarting Claude Desktop or writing manual stdin test harnesses.

MCPOp sits invisibly between your AI client and your MCP server, records the stdio JSON-RPC traffic, and flags these failures in real time. It does not currently block retries or intercept the call; the dashboard tells you what is happening so you can stop the loop yourself.

---

What MCPOp does

MCPOp is a single CGO-free Go binary. The proxy forwards each newline-delimited JSON-RPC line immediately, then parses it asynchronously for storage and heuristics.

• Loop Detection: Flags consecutive failed tool calls with identical arguments before they spiral into token-burning loops.
• Real-Time Schema Validation: Cross-checks argument payloads against the server's tools/list schema to immediately pi