Spanlens

Description

If your app calls an AI model, you eventually hit the same wall every team hits: you have no idea which feature is burning your API budget, why a specific agent run failed three steps in, or whether a user's private data accidentally ended up in a prompt log. Spanlens is built to answer exactly those questions — it sits between your code and the AI provider, watching every request go by, without you needing to rewrite how you call the model.

Technically, integration is a single baseURL swap that routes calls through Spanlens instead of directly to the provider, supporting 10+ providers including OpenAI, Anthropic, Gemini, Groq, Mistral, and xAI. From there it gives you per-model cost tracking with live spend analytics (and doesn't bill you again for cache hits), visualizes multi-step agent workflows as nested trace spans so you can see exactly where a chain of tool calls went wrong, supports OpenTelemetry (OTLP/HTTP) for plugging into existing observability stacks, and includes built-in eval grading you can wire into CI/CD to catch regressions before they ship. It also flags anomalies and potential PII in logged requests, though how reliable that PII detection is in practice isn't independently verified yet.