The platform

Per-request cost attribution, P99 latency tracking, and token throughput — across every provider

Framewren instruments your LLM API layer and normalizes cost, latency, and throughput data across every model provider into one dashboard. Not a tracing tool. Not an eval framework. LLM infrastructure observability, nothing else.

Start free trial Read the docs

Cost visibility

Per-model cost attribution, down to the request

Know exactly where every dollar of LLM spend is going. Break down costs by model, project, user, and time window with automatic anomaly flagging.

Cost breakdown  ·  June 2026 Budget: $5,000/mo
Month-to-date
$3,841
vs. last month
+18%
Projected EOm
$4,920
GPT-4o  $2,140
Claude 3.5  $980
Gemini Pro  $480
Mistral  $241

Latency tracking

P50 / P95 / P99 — charted over time

Surface latency regressions instantly. Know whether degradation is provider-wide or project-specific, and get alerted before your users notice.

Latency percentiles  ·  7-day view OpenAI GPT-4o
P50
420ms
P95
1,840ms
P99
3,120ms
regression
—— P50 —— P95 —— P99

Multi-provider

One dashboard across every provider you use

Framewren normalizes cost and latency metrics across OpenAI, Anthropic, Gemini, Mistral, Cohere, and self-hosted models into a single unified view.

Diagram showing Framewren connecting multiple LLM providers into a unified observability layer

Zero-prompt-logging architecture

We instrument the API layer, not the content. Your prompts and completions never leave your infrastructure. Framewren sees token counts, costs, and latencies only.

Automatic provider normalization

Different providers use different pricing units (tokens vs. characters vs. API units). Framewren normalizes everything to a common cost-per-request model you can compare directly.

Export & integrate

Export metrics as CSV, push to your data warehouse via webhook, or query the REST API directly. Framewren data fits into your existing engineering workflow.

Alerting

Smart alerts before the damage is done

Set budget thresholds, latency regression triggers, and throughput anomaly detectors. Alerts fire to email, Slack, or any webhook — your team decides when to respond.

Active alert  ·  Budget threshold High

GPT-4o daily budget threshold reached (85%)

Projected overage: $340 by end of day based on current call rate.

project: search-pipeline webhook: delivered ✓ slack: #eng-alerts

Security & data minimization

Observe the infrastructure layer. Never the content layer.

Framewren is designed to collect exactly the data needed for cost and latency analysis — and nothing else. Prompt text, completion content, and user data never leave your infrastructure. This is enforced by the SDK architecture, not by a privacy policy.

Zero prompt storage

Framewren never stores prompt text or completion responses. The SDK intercepts metadata only — token counts, model identifiers, latency measurements.

HTTPS-only, TLS 1.3

All data in transit uses TLS 1.3. Metric ingestion endpoints accept only authenticated requests. API keys are salted, hashed, and never stored in plaintext.

Row-level isolation

Each organization's data is isolated at the database row level. Multi-tenant architecture designed with principle-of-least-privilege access controls throughout.

Get started

One framewren.wrap() call and you're live

No infrastructure changes. No proxy deployment. No prompt routing layer. Wrap your existing LLM client and see cost and P99 data in the dashboard within 30 seconds of your first request.

Start free trial Read quickstart