Tokmeter
Token and cost observability for AI coding agents — nineteen agents, five surfaces, benchmarked against a 77 GB corpus.
- Parses local session logs from 19+ coding agents — Claude Code to Cursor to Zed
- Relay store — per-day immutable aggregates, ~170× smaller on disk
- Real benchmarks on a 77 GB, 319k-record corpus; 1–6 ms warm reads
- Cost report card — burn rate, compaction tax, cache-efficiency grade
If you run AI coding agents all day, you’re spending money you can’t see. Tokmeter reads the session logs those agents already write to disk and turns them into per-project, per-model, per-day truth — no telemetry, no cloud, no account.
Five surfaces off one core: CLI, Ink TUI, web dashboard, an MCP server with 24 tools, and a macOS menubar whose icon is a live health gauge.
The engineering write-up publishes its own misses — including a memory optimization whose early numbers didn’t survive honest measurement. That section stays in the README on purpose.