Spanlens

Spanlens is open-source LLM observability built for application developers. Drop in a one-line proxy or SDK swap and every OpenAI, Anthropic, or Gemini call your app makes is captured with full request and response body, model, tokens, cost, and latency.

Core features:

• Drop-in proxy. Replace api.openai.com with your Spanlens endpoint and every call is logged, no callback handlers or per-chain instrumentation needed.
• Agent tracing with Critical Path. Multi-step agents render as waterfall span trees with the longest dependency chain highlighted automatically.
• Prompt A/B with Welch t-test. Compare prompt variants side by side with real statistical significance on latency, cost, and error rate.
• Model savings recommender. "Swap these gpt-4o classification calls to gpt-4o-mini, $412 a month saved" with the evidence.
• Security built in. PII detection (SSN, credit card, IBAN, passport), prompt-injection patterns, and automatic API key masking before any row is persisted.
• 3-sigma anomaly detection on latency and cost against a rolling 7-day baseline.
• Versioned prompt library with diff, A/B rollout, and playground.
• OpenTelemetry (OTLP/HTTP) ingest for everything else.

Self-host the entire stack with one Docker Compose file. The whole repository is MIT licensed, no enterprise edition folder. Data stays in your infrastructure.

Free tier: 50,000 requests/month, 14-day retention, all core features. Pro from $29/month.

Positioned as a lightweight, developer-first alternative to Helicone (now in maintenance after Mintlify acquisition) and Langfuse (heavy, requires SDK wrapping every call).