Agent tooling · July 2026

LLM Cost Ledger

An MCP server that answers a question most people running CLI coding agents can't: what is this actually costing me? It reconciles real inference cost from local usage logs across Claude Code, Codex, OpenCode, and Amp.

Reading logs that already exist

Rather than instrument every tool separately, it leans on ccusage to read each CLI's local usage logs, then normalizes that output into a common set of usage records. An adapter layer absorbs the differences between tools; a pricing layer turns normalized usage into an actual dollar figure.

One MCP server, four CLI agents.Claude Code, Codex, OpenCode, and Amp usage reconciled into one cost model.
Terminal running the ccusage smoke check across five CLI agent sources, each returning OK with a day-row count
The adapter's smoke check, run live against real local usage logs on this machine.

Cutting scope to hit a deadline

The original plan included carbon and energy modeling alongside cost. Under a hard deadline, v1 scope narrowed to cost-only, with the reasoning for that cut written down in the repo rather than silently dropped, so the carbon-modeling path stays picked-up-able later instead of forgotten.

PythonMCPuvpytest