Ship agents you can actually debug.
You are writing the agent, picking the model, wiring the tools, and on the hook when it breaks at 2am. Trefur is the trace you wish you had the first time a customer pasted a screenshot at you.
What you are trying to do.
- →Get an agent from prototype to production without losing the thread on what it is doing.
- →Reproduce a failed run from a customer report — exact prompt, exact tool response, exact retry.
- →Catch a regression before customers do, when you swap models or change a prompt.
- →Keep cost-per-run visible while you iterate, not after the invoice arrives.
- →Use the framework you already chose, not the one your observability vendor prefers.
What hurts today.
Print statements and log greps
Your agent does five things in a single user turn — three LLM calls, two tool invocations, one retry. The terminal scrollback is a mess and the prod logs do not reconstruct the order.
LLM-only tooling misses the half that broke
Standard LLM observability shows you the prompt and the completion. It does not show you the tool that returned the wrong shape, the JSON the model failed to parse, or the silent retry that ate your budget.
Every framework speaks a different dialect
LangChain callbacks, CrewAI events, your custom orchestrator — three weeks of glue code just to get them on one timeline. You stop instrumenting and start guessing.
Cost shocks at the end of the month
The bill arrives, and you cannot answer which agent run drove the spike. By the time you find it, the prompt that caused it has shipped to a thousand users.
What Trefur gives you.
One trace per run
A single root span per agent run. Every LLM call, tool call, retry, and outcome hangs off it. Search by trace ID, agent name, or customer interaction.
Drop-in instrumentation
Auto-instrumentation for LangChain, CrewAI, AutoGen, Pydantic AI, MCP. One line at startup; existing agent code runs unchanged.
Cost and latency on every span
Tokens, latency, and dollars attributed to the step that emitted them. Roll up per agent, per build, per deployment.
Open standards, no lock-in
Trefur reads OpenTelemetry GenAI semantic conventions natively. Add it as a second OTel exporter — keep your existing pipeline.
Alerts that mean something
Route to Slack, PagerDuty, Teams, webhook. Alert on retry-rate spikes, cost-per-run drift, error-rate envelopes — not raw token counts.
Self-hostable collector
Single Go binary inside your network. Local batching, on-disk buffering, redact PII before anything leaves your perimeter.
A day in the life with Trefur.
- 09:14
Slack pings: a customer says the support agent forgot their order. You open Trefur, paste the trace ID from the support reply, see the agent retried a 429 from your order service, then gave up after three attempts.
- 09:22
You open the failing span and see the exact prompt, tool args, and response it captured, so you can reproduce the bug locally, bump the retry budget, and redeploy.
- 11:30
You swap the model on staging from gpt-4o to claude-sonnet for a cheaper draft. Trefur shows tokens dropped 22%, p95 latency dropped 410ms, retry rate held steady. You promote the change.
- 14:05
Cost-per-run alert fires for the writer agent. You drill in: the prompt picked up a long system message after a refactor. You trim it, redeploy, watch cost normalize in the live chart.
- 17:40
You head out. Oncall is covered by alerts on retry-rate spikes and error-rate envelopes — routed to PagerDuty with the trace already attached.
Get your first trace in under five minutes.
Free tier. No card required. Full feature access for individual developers.