Full visibility into every LLM call
Log calls, track versions, compare models, and optimize cost — all in three lines of code.
Log every LLM call
Latency, tokens, model, cost — captured automatically for every request. No agent frameworks required.
Prompt version tracking
Tag every call with a version string. See cost, latency, and token counts broken down per version.
Real-time cost tracking
Spend per prompt, per model, and projected monthly cost — updated live. Catch a prompt going rogue before your bill does.
Latency monitoring
Track p50/p95 latency per prompt and version. See the histogram, catch regressions, and get alerted when p95 goes red.
p50
420ms
p95
1.2s
Model comparison
Run the same prompt on Haiku, Sonnet, and Opus simultaneously. Compare output, latency, and cost side by side before you ship.
Cost intelligence
Automatic model downgrade suggestions, caching opportunities, and token reduction tips — with confidence ratings and savings estimates.
Anomaly detection
Cost spikes and slow calls are flagged automatically. Know within seconds when something breaks in production.
Hard cost caps
Set a monthly call limit and logging pauses when you hit it — your app keeps running, you just stop being tracked. No overages. Ever.
Up and running in 60 seconds
Works with any LLM provider — OpenAI, Anthropic, or any OpenAI-compatible API.
Install the SDK
npm install @stemma/sdk
Initialize with your project key
import { Stemma } from "@stemma/sdk";
const stemma = new Stemma({ projectKey: process.env.STEMMA_KEY });Wrap your LLM call
const result = await stemma.wrap({
promptId: "summarizer",
version: "v2.0",
call: () => anthropic.messages.create({ model, messages }),
});Know exactly which prompt version costs what
Tag every LLM call with a prompt ID and version string. Stemma groups metrics by version automatically — side-by-side latency, cost, p95, and token counts. No custom tooling.
- p50 / p95 / p99 latency per version
- Cost per call with monthly projection
- High-latency alerts and suggested experiments
Version Analysis · summarizer
Pick the right model before you ship
Run any prompt on Haiku, Sonnet, and Opus simultaneously. See outputs, latency, and cost side by side — so the choice is data, not instinct.
- Compare up to 5 Claude models at once
- Fastest and cheapest model highlighted
- Token breakdown per model
Model Comparison · 3 models
Haiku 4.5
Fastest · CheapestLatency
1.2s
Cost
$0.0012
Sonnet 4.6
Latency
4.1s
Cost
$0.0098
Opus 4.6
Latency
8.7s
Cost
$0.082
Automatic optimization recommendations
Stemma analyzes your usage and surfaces actionable suggestions: model downgrades, token reduction opportunities, and caching wins — each with a confidence rating based on your actual call history.
- Savings range with ±20% confidence interval
- Current vs. optimized annual cost
- Stacked model usage bar for visual comparison
Cost Intelligence
⚡ Switch to Haiku 4.5
4 calls on Sonnet 4 · 3–5× faster
$0.11–$0.17
/month
Model Usage
→ Haiku 4.5 saves $0.16/mo
Replay any production call in the browser
Every logged call can be replayed directly in the Stemma playground. Edit the prompt, swap the model, and compare outputs — without touching your production code.
- One-click replay from any log entry
- Edit prompt and re-run live
- Latency and cost shown per run
Playground · summarizer v1.0.0
system
You are a concise summarizer. Return 3 bullet points.
user
Summarize this lease agreement: The tenant agrees to pay…
assistant · 4.2s · $0.008
• Monthly rent: $2,400 due on the 1st
• 12-month term, no early exit
• Pets allowed with $500 deposit