# Helicone Helicone is open-source LLM observability built around a gateway proxy. You change one line — point your OpenAI/Anthropic base URL at Helicone — and you get logs, traces, costs, latency, caching, and rate limits without touching SDK code. There's also an async logging path if you don't want a proxy in the request path. The "1-line install" angle is the differentiator. [[LangSmith]] and [[Langfuse]] need you to wrap calls or set up tracers; Helicone catches them at the wire. ## What it gives you - **Proxy or async logging** — proxy for the easiest setup; async for teams that won't accept a third party in the hot path. - **Cost tracking** — per request, per user, per model, per session. The dashboard the finance person actually wants to look at. - **Caching** — semantic and exact caching at the gateway. Drop production cost by skipping redundant calls. - **Rate limits** — per user or API key, enforced at the proxy. Useful for multi-tenant apps where one user can't burn the whole budget. - **Prompt management** — versioned prompts and a playground; newer surface than the observability core. - **Sessions, agents, evals** — caught up to the broader feature set over time; less mature than [[Langfuse]] / [[LangSmith]] on evals specifically. ## When I'd reach for it - You want LLM observability **today**, with one URL change, before you commit to a full SDK integration. - Cost control is the actual driver, not eval workflows. - You need caching or per-user rate limits at the gateway layer. ## Trade-offs - **Proxy in the request path** — added latency hop and a new failure point. The async mode avoids this but adds complexity. - **Eval / dataset tooling is less rich** than Langfuse or LangSmith. If structured evals are your priority, those are stronger. - **Self-hosting is supported** but the cloud product is the polished path; pick what you actually need. ## Alternatives - [[LangSmith]] — closed source, polished, eval-heavy, LangChain-aligned. - [[Langfuse]] — open source, self-hostable, deepest prompt + eval tooling of the three. - [[Edgee]] — gateway pattern too, but optimized for agent-to-LLM routing and token compression rather than full observability UI. - [[LiteLLM Proxy Configuration]] — proxy for unifying provider APIs; pair with one of these for observability. ## References - Website: https://www.helicone.ai/ - GitHub: https://github.com/Helicone/helicone - Docs: https://docs.helicone.ai/ ## Related - [[LangSmith]] - [[Langfuse]] - [[LLM Monitoring]] - [[AI Observability]] - [[Edgee]] - [[LiteLLM]] - [[LiteLLM Proxy Configuration]] - [[MLflow]]