mcpspan is a self-hosted analytics tool built specifically for MCP servers, and the latest release is all/v0.7.0. An MCP server is a server that provides tools, resources, and prompts to agent clients (Claude, Codex, Cursor, and the like). mcpspan records which client called which item, how slowly it was called, and why it failed. There is no public description of what all/v0.7.0 changed, so only the features documented in the repository are covered here.
The gap this tool fills is the observability gap in agent operations. Teams running MCP servers have a hard time knowing which tools are slow, which calls fail, and what keeps repeating. mcpspan records usage and response times for tool calls, resource reads, and prompt requests, lets you inspect call outcomes, error types, and repeated calls within a session, and compare differences across server versions. It also tracks how many tokens the tool definitions shown to the agent consume. The primary language is TypeScript, and the license is MIT.
The implementation is split across several languages. It ships connector tools for TypeScript, Python, Go, C#, Java, Rust, Ruby, and PHP, together with a shared conformance test suite, so you can check how servers written in different languages behave differently in the same way. Data stays inside the servers you configure, and CSV and NDJSON export plus OpenTelemetry delivery are supported.
That said, the thin parts of the evidence are clear. The repository has 18 stars, and in the research record the momentum score is measured low at 1.2, which puts it closer to an early-stage project than a widely used tool. There is no published baseline for what latency and token consumption to expect, or for whether servers in different languages behave at the same level. Because the data stays inside the server, there is little to hand off externally, but that also means there is nothing to compare against. For teams already running MCP servers, it is a reasonable starting point for finding out what is slow and what repeats.