AI FinOps

Calculator

Compute the cost of trusted work

Two calculators, one discipline. For development, compute cost per accepted change — the cost of producing software that stayed in production. For agents, compute cost per accepted action — the cost of agent work that stayed accepted. Both run entirely in your browser; nothing is sent to a server.

Cost per accepted change — for one measurement window (a sprint, a month, a quarter).

Cost per accepted change formula, rendered as a diagram COST PER ACCEPTED CHANGE model cost · infrastructure · engineering time · review · rework accepted change units ( reached production and stayed there ) $ AC THE COST OF DELIVERED SOFTWARE THAT STAYED DELIVERED
Development inputs
LLM / API spend attributable to the changes produced.
Compute, storage, observability, tooling overhead.
Spec / prompt / integrate / steer, converted to currency.
Time spent reviewing / gating AI work, converted to currency.
Cost of fixing or reverting changes that did not stay in production.
Whole-number count of changes that reached production and stayed there.
Cost per accepted change —
What this measures →

Use the library

Both calculators are pure TypeScript functions — import them into your dashboards, FinOps pipelines, or CI jobs: 

import { costPerAcceptedChange, costPerAcceptedAction } from 'cost-per-accepted-change';

// Development: software that stayed in production
costPerAcceptedChange({ modelCost: 1200, infraCost: 400, engineeringTime: 18000,
  reviewCost: 6000, reworkCost: 2400, acceptedChanges: 42 }).value;   // 666.67

// Agents: actions that stayed accepted (failureImpactCost = escalations,
// refunds, lost sales — the downstream consequence of actions that failed)
costPerAcceptedAction({ inferenceCost: 3000, toolCost: 1200, infraCost: 800,
  oversightCost: 9000, remediationCost: 4000, failedRunCost: 1000,
  failureImpactCost: 12000, acceptedActions: 5000 }).value;             // 6.2

Source: github.com/brennhill/cost-per-accepted-change. MIT-licensed.

Use the CLI

For one-shot calculations and full repo audits, install the CLI. cpac audit walks a GitHub repo over a date window, detects reverts, applies the 500-LOC normalization, and emits the accepted-change-unit count: 

# Pure math
npx cost-per-accepted-change-cli cpac calc \
  --modelCost 1200 --infraCost 400 --engineeringTime 18000 \
  --reviewCost 6000 --reworkCost 2400 --acceptedChanges 42

# Audit a real repo window for the denominator
npx cost-per-accepted-change-cli cpac audit \
  --repo owner/name --since 2026-04-01 --until 2026-04-30

Requires the GitHub CLI on your PATH and an authenticated session.

Use the MCP server

The same core is exposed as an MCP stdio server (cpac-mcp) so AI agents in Claude Code, Cursor, and other MCP-aware clients can compute the metric inside a conversation: 

// Add to your MCP client configuration
{
  "mcpServers": {
    "cpac": {
      "command": "npx",
      "args": ["-y", "cost-per-accepted-change-cli", "cpac-mcp"]
    }
  }
}

The skill manifest is published at /.well-known/agent-skills/cost-per-accepted-change.md and indexed at /.well-known/agent-skills/index.json. Source: tools/cpac-cli.