Ratchet — agent friction, measured

Ratchet is a machine-local observer for AI coding agents. It ingests your own agent transcripts into a local SQLite store, prices them with cache-split accounting, detects recurring shapes of friction, and ranks the hotspots by what they actually cost — so the next agent to open your project can query what went wrong last time instead of rediscovering it.

• Six friction signatures — WAIT, REPEAT, PAGING, REPHRASE, SERIAL, CEREMONY — each a shape seen repeatedly in real transcripts
• Cache-split pricing — cache reads are not input tokens, and unknown cost is NULL, never zero
• Honest caveats computed every run — approval wait is labelled an upper bound, never machine time
• Everything stays on the machine — no service, no account, no telemetry; only derived findings are portable
• MCP-native — one-call orientation for agents, plus a session-start packet of the project's measured traps
• Ships with atlas — ask a word, get a verified file:line pointer in whatever project you stand in

Ratchet is live and free at ratchet.daystra.com. Get in touch about how we use it.