AI observability

See what your models are actually doing.

Irida traces every prompt, tool call, and token your stack produces. Latency, cost, and eval quality on one screen, streaming live.

Start tracing Watch a trace

OTLP native 2ms SDK overhead 30 second setup

Scroll

Instrumented at

Nightjar Ferrite Labs Parsec Cobalt Loop OKTAV Heliotrope

01 Observe

Every call, on the record.

Two lines to instrument. After that, Irida catches everything your agents do and shows you the parts that matter.

Traces stream in as they happen

No sampling lag, no batch delay. Spans land in your console under a second after they fire in prod.

Cost, per span

Know exactly which prompt burns the budget, down to the retry.

openai.chat$18.40 / day

Evals that watch for drift

Score every response against your rubric. Catch quality slipping before your users do.

groundedness · 7d0.94 → 0.71

Prompt diffs

Every version tracked. See what changed and what it did to quality.

Alerts with teeth

Thresholds on latency, cost, and eval score. Routed before the postmortem writes itself.

Session replay

Step through an agent run message by message, exactly as it happened.

msg 1msg 14 of 224.2s

02 Inspect

One trace tells the whole story.

Every span in the run, timed and priced. The waterfall shows you where the four seconds went.

traces /tr_9f3a2c ok retry x2 4.21s · 7 spans

agent.run4.21s

├guard.pii_scan96ms

├retrieve.chunks540ms

├openai.chat · gpt-4o1.58s

├tool.web_search660ms

├openai.chat · gpt-4o902ms

└eval.groundedness178ms

{
  "trace_id": "tr_9f3a2c",
  "env": "prod", "region": "us-east",
  "duration_ms": 4212, "status": "ok",
  "tokens": { "in": 5214, "out": 1088 },
  "cost_usd": 0.0431,
  "evals": { "groundedness": 0.94 },
  "spans": [
    { "op": "agent.run", "ms": 4212, "children": 6 },
    { "op": "guard.pii_scan", "ms": 96 },
    { "op": "retrieve.chunks", "ms": 540, "k": 12 },
    { "op": "openai.chat", "model": "gpt-4o", "ms": 1580 },
    { "op": "tool.web_search", "ms": 660, "retries": 2 },
    { "op": "openai.chat", "model": "gpt-4o", "ms": 902 },
    { "op": "eval.groundedness", "ms": 178, "score": 0.94 }
  ]
}

Irida flagged this run. Retry loop in tool.web_search added 660ms. p95 latency is up 34% since prompt v41 shipped. See the diff

2ms

SDK overhead, p99

1.2B

Spans traced per day

99.99%

Uptime SLA

<1s

Ingest to screen

03 Pricing

Pay for what you trace.

Spans are the only meter. No seats math, no ingest surprises, no annual call with sales to see a number.

Trace

For side projects

$0forever

50k spans a month
7 day retention
Live trace stream
1 project, 2 seats

Start free

Most teams pick this

Lab

For teams shipping to prod

$79/mo

5M spans a month
30 day retention
Evals, diffs, and alerts
Session replay
Unlimited projects, 10 seats

Start 14 day trial

Observatory

For platform teams

Custom

Unlimited spans
13 month retention
Self-host or private cloud
SSO, audit log, 99.99% SLA

Talk to us

Stop guessing.
Start seeing.

Your models are talking. Irida writes it all down.