AI Economics

What ai-economics does

The ai-economics MCP server exposes twelve focused calculators for assessing the operational and economic consequences of AI systems. It covers token pricing across GPT, Claude, Gemini, and DeepSeek; context-window sizing; hourly agent costs; model-tier routing; query energy use; and organization-wide token consumption measured in money, electricity, and carbon dioxide.

The calculators also address whether AI work can be trusted or supervised. They estimate energy per verified task, autonomy after evidence and independent validation are required, the time revoked credentials may remain usable, the number of agents a human review team can handle, and the accumulated cost of unverified work. A separate calculator estimates how long a humanoid robot can operate on a charge.

How it works

The MCP process wraps the free piszczek.pl/tools API. Each tool accepts optional parameters; when values are omitted, the server uses defaults corresponding to the interactive calculators.

Responses include four fields: a calculated result, the formula used, an interpretation, and a cite_as sentence suitable for quoting. This makes the output useful both for exploratory questions and for documents that need to explain how a figure was derived. The project is listed in the official MCP Registry as pl.piszczek/ai-economics.

Setup and configuration

For Claude Desktop or Claude Code, run the package through npx:

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

Claude Code also supports claude mcp add ai-economics -- npx -y @michalpiszczek/ai-economics-mcp. The package is plain JavaScript and does not need a compilation step. For local stdio execution, it can be built into a Docker image and run interactively without exposing a port.

AI_ECONOMICS_API is optional. Set it to replace the default https://piszczek.pl/tools/api endpoint, such as when pointing the client at a self-hosted instance. No credentials are required by the documented setup.

Tools and capabilities

The ai-economics MCP server provides these tools:

• token_cost compares monthly costs for one million tokens across listed model families.
• agent_hour calculates a fully loaded hourly agent cost that includes verification.
• model_routing estimates annual savings from sending work to a cheaper model tier.
• llm_energy and token_burn estimate electricity use, with token_burn also covering financial cost and CO₂.
• joules_per_verified_task compares energy efficiency for tasks that pass verification.
• proof_adjusted_autonomy, verification_bottleneck, and proof_debt model evidence, human review capacity, and the cost of unchecked work.
• humanoid_energy estimates robot runtime per charge, while revocation_exposure estimates how long revoked authority may continue working.

Limitations and notes

Results depend on the assumptions supplied to each tool. The server connects to the configured calculator API rather than storing calculation inputs locally. Concepts and calculators are licensed separately from the MIT-licensed code: the README identifies the concepts and calculators as CC BY 4.0.

The ai-economics MCP server is best suited to budgeting, comparative analysis, and discussions of verification or operational exposure. It is not documented as a monitoring system, an access-control mechanism, or a replacement for independent energy, cost, or security measurements.