# Metrx

> Metrx heads every AI workload toward its best proven configuration. It finds cheaper and better configurations, proves them against a randomized holdout, and switches on your say-so — as models and prices change.

## What Metrx Does

- Generates candidate configurations per workload: model, prompt, parameters, routing
- Proves them with pre-registered randomized trials on your traffic, once it has evidence
- Judges every candidate against an acceptance contract you approve, not a generic score
- Can put winners live under the authority you define, per change class
- Watches live configurations for drift and re-verifies incumbents on model releases
- Rolls back on contract violation and records the event
- Keeps a full evidence ledger — wins, losses, and rollbacks alike
- Tracks per-agent, per-model LLM cost and token use across providers

## What Metrx Does NOT Do

- Metrx does NOT store prompt or completion content — only metadata, cost signals, and the outcome signals you choose to send
- Metrx is NOT a prompt management tool — use a prompt registry for versioning
- Metrx is NOT an agent hosting platform — it works with your existing stack
- Metrx does NOT claim a population benchmark is your result — every number carries its provenance

## Procedural Integrity

Metrx optimizes and also measures, so the trust claim is about procedure, not about neutrality:

1. Candidate generation is separated from evaluation
2. You approve the objectives and the acceptance contract; optimization happens inside them
3. Treatment is randomized, so measured effects are causal rather than before/after coincidence
4. Negative and inconclusive results are preserved — never retried until positive
5. Every promotion carries a reproducible evidence record with provenance and exclusions

## Provenance Labels

Every number on every Metrx surface carries one of these:

- Measured — directly observed from production data
- Attributed — a modeled estimate with stated confidence
- Verified — produced by a pre-registered randomized trial whose negatives are preserved
- Estimate — a projection, labeled as such
- Synthetic — demo or illustrative data, never implied real

## For AI Agents (MCP Server)

Install (stdio): `npx @metrxbot/mcp-server`
Remote (HTTP): https://metrxbot.com/api/mcp

The authoritative, machine-readable tool registry — names, descriptions, domains and count — is published at https://metrxbot.com/.well-known/agent-card.json. Read it there rather than from prose, so an agent never works from a stale list.

## MCP Configuration

```json
{
  "mcpServers": {
    "metrx": {
      "command": "npx",
      "args": ["@metrxbot/mcp-server"],
      "env": { "METRX_API_KEY": "sk_metrx_..." }
    }
  }
}
```

## For Developers (REST API & SDKs)

- REST API: https://metrxbot.com/api/v1/
- OpenTelemetry: send OTLP traces to https://gateway.metrxbot.com/v1/traces
- Python SDK: `pip install metrxbot`
- TypeScript SDK: `npm install @metrxbot/sdk`

## Quick Start

1. Sign up at https://metrxbot.com (free, no credit card)
2. Get an API key from Settings → Security
3. Install the MCP server or an SDK
4. See labeled cost data in minutes; candidates follow once a workload has traffic

## Links

- Documentation: https://docs.metrxbot.com
- API Reference: https://docs.metrxbot.com/api-reference
- Agent Quickstart: https://metrxbot.com/agent-quickstart
- Machine-readable tool registry: https://metrxbot.com/.well-known/agent-card.json
- Machine-readable pricing (JSON): https://metrxbot.com/api/pricing
- Pricing: https://metrxbot.com/pricing
- Full documentation: https://metrxbot.com/llms-full.txt