LLM reliability · SRE for agents
Your AI agents fail quietly.
Puse makes it loud.
Reliability monitoring for LLM apps and agents: every model call traced — latency, tokens, cost, errors — with alerts when error rates climb or spend spikes. SRE-grade observability, sized for teams shipping with AI today.
Instrument once, see everything
import Anthropic from "@anthropic-ai/sdk";
import { createPuse } from "@puse/sdk";
const puse = createPuse({ apiKey: "pu-..." });
const anthropic = puse.wrapAnthropic(new Anthropic());
// use the client exactly as before — streaming included.
// every call is traced: model, tokens, cost, latency, errors.
const stream = await anthropic.messages.create({
model: "claude-sonnet-4-20250514",
max_tokens: 1024,
messages: [{ role: "user", content: "…" }],
stream: true,
});Built for reliability, not vanity metrics
The pulse of your LLM apps — trace every call, catch regressions, page the right person.
One-line auto-instrumentation
Wrap once. Trace everything.
Wrap your OpenAI or Anthropic client once. Every call — including streams — is traced with model, tokens, latency, and cost. No proxy, no gateway in your request path.
Streaming-aware cost tracking
Know spend mid-stream.
Token usage is captured from streamed responses as they flow, priced against a per-model table you control. Know what every agent run cost, even mid-stream failures.
Health dashboard
See the spike first.
Requests, error rate, latency percentiles, and spend — per project, charted over 24 hours or 7 days. See the spike before your users do.
Alerts that page you, not spam you
Pager-style reliability.
Error-rate and cost-spike detection with tunable thresholds, delivered to Slack. Built for reliability, not vanity metrics.
Errors with context
3am becomes actionable.
Failed calls keep their error message, model, and latency — so 'it broke at 3am' becomes 'rate-limited on claude-sonnet for 12 minutes'.
Self-hostable, tiny footprint
Your traces stay yours.
A Fastify ingest service, a Next.js dashboard, Postgres. Runs on one small VPS. Your traces never leave your infrastructure.
Stop finding out from your users
The first sign of a broken agent shouldn't be a churned customer. Put a pulse on every model call today.