AI Economics README

ai-economics-mcp

MCP server for AI economics. Gives Claude, Cursor and any MCP client 12 calculators for the questions that decide AI budgets: what tokens cost, what queries burn, and whether agent work can be proven.

Listed in the official MCP Registry as pl.piszczek/ai-economics (domain-verified). Wraps the free piszczek.pl/tools API — no key, no sign-up, stateless (inputs are never stored). The math is the same as the interactive calculators.

Install

Claude Desktop / Claude Code (claude_desktop_config.json / .mcp.json):

config.json

{
  "mcpServers": {
    "ai-economics": {
      "command": "npx",
      "args": ["-y", "@michalpiszczek/ai-economics-mcp"]
    }
  }
}

Claude Code one-liner:

Terminal

claude mcp add ai-economics -- npx -y @michalpiszczek/ai-economics-mcp

Tools

Tool	Answers
`token_cost`	What do 1M tokens cost per month across GPT, Claude, Gemini, DeepSeek?
`context_window`	How many tokens is 50 pages, does it fit, what does carrying it cost?
`agent_hour`	What does one hour of an AI agent cost, fully loaded with verification?
`model_routing`	How much does routing to a cheaper tier save per year?
`llm_energy`	How much electricity does an AI query use?
`joules_per_verified_task`	Which model is most efficient per task that actually passes?
`token_burn`	What does org-wide token burn cost in money, kWh, CO₂?
`humanoid_energy`	How long can a humanoid robot run per charge?
`proof_adjusted_autonomy`	How autonomous is a "90% autonomous" agent once proof is required?
`revocation_exposure`	How long does a revoked token keep working across your stack?
`verification_bottleneck`	How many AI agents can a human review team absorb?
`proof_debt`	What does unverified AI work cost over time?

All parameters are optional; defaults mirror the interactive calculators. Every response includes result, formula, interpretation and a ready-to-quote cite_as sentence.

Example

You: how autonomous is our agent really? it completes 90% alone, evidence for 95%, 80% independently validated, 90% on time

Claude (calls proof_adjusted_autonomy): PAA = 61.6% — supervised autonomy. The gap vs the claimed 90% is 28.4 pp, and the biggest lever is independent validation…

Configuration

Env var	Default	Purpose
`AI_ECONOMICS_API`	`https://piszczek.pl/tools/api`	Point at a self-hosted instance

Glama builds and releases

Glama generates its own Dockerfile from the build configuration. Use these settings:

Build steps: ["npm ci --omit=dev"]
CMD arguments: ["mcp-proxy", "--", "node", "index.js"]
Environment variables: optional AI_ECONOMICS_API; no credentials required.

This is plain JavaScript: there is no build script or compilation step. Do not use pnpm run build or npm run build. After syncing the repository, build the selected commit and publish a Glama release from the successful test. A GitHub or npm release alone does not trigger Glama's quality evaluation.

For local stdio use, the repository Dockerfile needs no proxy or exposed ports:

Terminal

docker build -t ai-economics-mcp .
docker run -i --rm ai-economics-mcp

Concepts behind the tools

Joule Wars — the AI race for energy efficiency
Proof-Adjusted Autonomy — the metric of proven agent work
Revocation Exposure — how long revoked authority keeps working

Citing a result

Every tool response carries a cite_as sentence written to be quoted verbatim. For a document that has to survive review, DOIs and BibTeX for this server, the calculators and the three concepts they instrument are at piszczek.pl/cite.

License

MIT. Concepts and calculators by Michał Piszczek (CC BY 4.0 — attribution appreciated).