// the find
Helicone/helicone
🧊 Open source LLM observability platform. One line of code to monitor, evaluate, and experiment. YC W23 🍓
Helicone is an LLM observability platform that works by proxying your API calls through their infrastructure — change one line (the baseURL), and you get request logging, cost tracking, and latency metrics across 100+ providers. It's for teams running LLM workloads in production who want visibility without rewriting their client code.
- The proxy integration pattern is genuinely clever: zero changes to business logic, just a baseURL swap. Works with any OpenAI-compatible client without SDK coupling.
- ClickHouse for analytics storage is the right call — Postgres would choke on high-volume request logs at any meaningful scale. The architecture shows they've actually thought about the data model.
- Self-hosting path exists and is realistic: docker-compose gets you a working stack, and the Helm chart means you can run it in k8s without rolling your own configs.
- The open-source LLM cost database (300+ models) is a genuinely useful artifact that stands on its own, separate from the platform.
- The Cloudflare Worker is in the hot path of every LLM request. When Helicone's worker has an incident, your app's LLM calls fail or degrade — you're adding a third-party dependency to a latency-sensitive code path.
- Self-hosting 'simple' requires Supabase + ClickHouse + Minio + Cloudflare Workers + the Jawn Express server — five separate services to operate. The docker-compose hides the operational complexity, not eliminates it.
- All your prompts and responses flow through their infrastructure in the hosted version. For any application with PII or sensitive data, that's a real problem the SOC 2 badge doesn't fully address.
- The codebase is clearly optimized around their SaaS product first, open source second. The Cloudflare Worker for proxy logging means the self-hosted path requires a Cloudflare account anyway, which limits true infrastructure independence.