Overview

Metergraph is a continuously learning system for your AI agent workflows. It cuts LLM costs and improves quality, and it keeps learning from your traffic as it arrives, so when your prompts or the available models change, you can find the next improvement.

It reads your LLM traces, shows where you are overspending or losing quality, recommends specific fixes, and helps you ship them. Every recommendation comes with its evidence: the real requests it was tested on, what each option answered, and how it was graded.

It also helps you build evals for your system from your own production traces, and gives you tools to experiment with changes and validate them before they ship.

How it works

1. Bring your tracesImport from Langfuse, LangSmith, Braintrust, Phoenix, Portkey, or OpenAI, or add the SDK in a few minutes.

2. See where cost and quality goSpend, calls, errors, and latency by agent workflow, model, and provider.

3. Get specific fixesEach one tested against your own traffic, with projected monthly savings and the effect on quality.

4. Ship themApprove a fix and MetergraphBot opens a pull request, or your coding agent applies it over MCP.

What it finds

Metergraph looks for changes that lower cost without lowering quality. Typical findings:

• A cheaper model with the same or better quality. Metergraph replays your real requests against candidate models and grades each answer against what your application produced.
• Prompt caching you are not using. Long system prompts and tool definitions sent on every call can be cached by the provider for a fraction of the input price.
• Inefficient context. Prompts that carry more context than the task needs, such as long histories or whole documents where an excerpt would do.
• Expensive tools. A search setup that costs more than the answer needs, and a cheaper search stack that answers just as well.

These are the most common findings, not the full list. Metergraph adds new kinds of finding over time.

Read how analysis works.

For developers

Tools for the engineers who own the workflow.

Analyze costBreak spend down by agent workflow, model, and provider, and compare traces side by side.

Build evalsTurn production traces into datasets and grading suites, or reuse your Langfuse judges.

Compare models on your trafficRun an analysis and see quality, cost, and latency for each candidate.

Review the evidenceSee the tested cases, the answers each option gave, and how they were graded.

Work from your coding agentQuery spend, start analyses, and apply fixes from Claude Code, Codex, or Cursor.

Get started

Import existing tracesAlready on Langfuse, LangSmith, Braintrust, Phoenix, Portkey, or OpenAI? Start here, with no code changes.

Send your first traceAdd the Python or TypeScript SDK to one real model call.

Run your first analysisTest lower-cost options against the work your application already does.

Give each agent workflow a name early, so Metergraph groups and analyzes its traffic together. Metergraph calls that name a route: name your first route.

Free and paid plans

Two ways to start free:

• Metergraph Cloud, Free plan. 100K calls a month with every feature: evals, recommendations, alerts, datasets, and MCP access.
• The open source server. An Apache-2.0 spend dashboard you run yourself. It discards prompt and response content at ingest and covers spend and pricing, not analysis, evals, or alerts.

Growth is the same product with a larger monthly allowance. Scale runs Metergraph in your VPC or on-prem, with SSO and managed optimization: our team runs the analysis and ships the pull requests with you.

Sign up free · Compare plans · Book a call

Your data and keys

Metergraph does not sit in your request path. The SDK records calls in the background, your provider keys stay in your application, and responses reach your code unchanged. You choose whether prompt and response content is stored.

Content and privacy · Choosing a deployment