Spanly

Spanly solves a real, growing problem: MCP servers are becoming a critical part of AI infrastructure, but standard APMs don't speak JSON-RPC. They see HTTP spans and stack traces, but not the tool call arguments, prompt payloads, or resource URIs that make MCP telemetry useful. Spanly fills that gap cleanly. We'd reach for this when your MCP server is in production and you need to debug a slow tool call or a recurring error without grepping logs. The CLI wraps any server (stdio or HTTP) with zero code changes, and the dashboard surfaces the metrics that matter: error rates, percentiles by tool and client, and full payloads for every request. The protocol-level capture means you see the exact argument that caused an invalid params error—something an APM's generic span can't show. Where it bites: Spanly is strictly MCP. If you need full-stack APM (infrastructure, databases, HTTP routes), you keep your existing tool. The pricing model—$6/100k requests above 100k—can add up fast at scale. For a busy server doing millions of requests, the Business tier at $210/mo is reasonable, but overage costs need monitoring. Compared to the closest alternative—just relying on your APM's MCP instrumentation or rolling your own logging—Spanly wins on speed of setup and depth of insight. Sentry and New Relic now offer MCP spans, but only for Node and Python, require code changes, and miss full payloads. Spanly works with any language and captures everything. In practice, we see this for: platform teams shipping an MCP server as part of a SaaS product; engineers debugging agent loops that hammer a tool with bad inputs; and teams that need to comply with EU AI Act traceability requirements (the blog post on that is a good read). Not ideal for hobby projects (<100k requests/month are