Debug AI agents without the guesswork.
When the output is wrong, the question is which step caused it. ObserveAI records every span and the prompt that was sent, so you can answer that without rerunning the pipeline.
Agent execution trace
Every step of the run, in the order it happened. Select a span to see its inputs and the response it produced.
- Chain
- Tool
- Retriever
- LLM
Trace ID: trc_8f92a4b1 · Latency: 4.2s
Daily spend and cost anomalies
Daily spend by model. A prompt that grew by two paragraphs shows up here weeks before it reaches the invoice.
- GPT-4o
- Claude 3.5
- Embeddings
Features
Two levels: see what your AI is doing, then set the rules for what it can do.
Observe
See what every model call does and what it costs.
Prompt debugging
Inspect exactly what goes in and out of your models. View raw prompts, model responses, latency, and token counts side-by-side to troubleshoot issues instantly.
Anomaly alerts
Get notified immediately via Slack, PagerDuty, or webhooks if your AI spend spikes or performance drops. Catch anomalies today, not at the end of the billing cycle.
Unit economics
Stop guessing where the money goes. Tag requests to break down your AI spend by team, tenant, or customer to clearly understand your true cost-to-serve.
Govern
Set the spending rules, keep an auditable record, and bring the cost down.
RBAC and auditing
Control exactly who can access which models. Maintain a detailed, exportable audit trail showing who made requests, when they happened, and what data left your network.
Policy enforcement
Prevent runaway costs automatically. Set strict budget caps and data rules at the proxy layer so over-quota requests are blocked before they ever hit the vendor.
Chargeback and optimisation
Accurately bill internal cost centres for their exact AI usage. Easily spot opportunities to lower costs by switching to cheaper models or optimising prompts.
Set up in an afternoon.
Point your existing model calls at ObserveAI and the dashboards fill in. There is no rewrite and no migration, and your teams keep building the way they do now.
import { ObserveOTel } from 'observeai';
import { OpenAI } from 'openai';
// 1. Your existing client, unchanged
const client = new OpenAI();
// 2. Configure the tracer
const tracer = new ObserveOTel({
tenantId: 'marketing',
capturePrompts: true,
});
// 3. Instrument the client
tracer.instrument(client);Drop-in SDK or proxy
Change one base URL to route through the edge proxy, or import the SDK to capture the reasoning spans a proxy can't see.
Vendor agnostic
OpenAI, Anthropic, Gemini, Llama and the common orchestration frameworks. Swap the model underneath and the history stays comparable.
Export anywhere
Spans are OpenTelemetry. Send them to Datadog, New Relic or Splunk and read AI cost next to the rest of your APM data.
Request early access
Tell us what you're running and we'll come back with a date. We onboard a small number of teams at a time so each gateway integration gets the attention it needs.