Long, multi-agent sessions are a black box. Sift reads your local session history and shows you the failure loops, token burn, file thrash, and runaway agents — so you can work faster and cheaper.
Token counters tell you how much. Sift tells you what went wrong — and what to do about it.
Real-dollar (API-equivalent) cost per model, project, tool, agent and MCP — deduplicated, so totals are honest. On a subscription, it’s the value you extracted.
Failure loops, thrashed files, stalls and edit-first churn — rolled into a single 0–100 Waste Score so the messiest sessions rise to the top.
Reconstructs the agent tree — per-helper cost, errors and loops — plus a clean diff of every file a session touched, no git required.
Watches your active sessions and pings you the moment one falls into a loop or stalls — catch the runaway before it burns an hour.
No setup, no config, no account. Point it at your Mac and go.
Sift reads your local Claude Code history (read-only). One click indexes thousands of sessions in seconds.
Each session gets a plain-English verdict, a Waste Score, and the exact moments it looped, thrashed or stalled.
Concrete suggestions for next time — and live alerts so the next runaway never gets far.
Privacy isn’t a feature bolted on — it’s the architecture. Sift is the local alternative to hosted session analyzers.
Reads ~/.claude on your Mac and never modifies it. The index is a disposable local cache.
Zero analytics. Zero network calls during normal use. Nothing to sign up for.
API keys and tokens are masked at the moment of parsing — before they reach the index, the UI, or any export.
Subscription fatigue is real. Buy it once, own it.