/// CLAUDE CODE · ANALYTICS
Sift
See what your Claude Code sessions actually did — and what they wasted.
/// A look inside
Your sessions at a glance
Sift opens on your most wasteful session, your total tokens, and what it all cost — so you instantly know where to look first.
A plain-English diagnosis
Every session gets a verdict and a 0–100 Waste Score, with the exact failure loops, thrashed files, and stalls that cost you — plus concrete suggestions to fix them next time.
See exactly what changed
A clean list of every file a session created, edited, or deleted — no git archaeology required.
Where your tokens actually go
Real-dollar cost broken down by model, project, and day. (API-equivalent — on a subscription it's the value you extracted, not your bill.)
Built for everyone
Plain-language explanations of every metric, with examples. You don't need to be technical to know what your sessions are doing.
A local-first Mac app that reads your Claude Code session history and tells you, in plain terms, what each session actually did and where it wasted effort — failure loops, token burn, file thrash, and runaway agents — without a single byte ever leaving your machine.
Power users run long, multi-agent Claude Code sessions with almost no visibility into their health. You can feel when one goes sideways — the same command failing over and over, a file rewritten fourteen times, an hour gone — but afterward it's a black box. Token counters tell you how much; nobody tells you what went wrong. Sift does.
What it does
Point Sift at your Mac and it indexes thousands of past sessions in seconds (read-only). For each one it produces:
- A plain-English verdict. "A lot of wasted effort — got stuck repeating the same failing test nine times, and re-edited auth.ts fourteen times." No jargon you weren't given.
- A Waste Score (0–100). Failure loops, file thrash, stalls, and edit-first churn rolled into one number, so the worst sessions sort straight to the top.
- Honest cost. Real-dollar (API-equivalent) spend per model, tool, agent, and MCP — de-duplicated so totals aren't inflated. On a subscription, it's the value you extracted, not a bill.
- The multi-agent tree. Which helper agents ran, what each cost, and whether one got stuck.
- What changed. A clean diff of every file the session created, edited, or deleted — no git required.
- Suggestions. Concrete advice for next time: when to break a loop, when to read before editing, when to split a session.
Catch the runaway live
Live mode watches your active sessions and fires a desktop notification the moment one falls into a failure loop or stalls — so you catch the runaway before it burns an hour, instead of finding out after.
Private by design
This is the whole point. Sift runs 100% on your Mac. Your session history is never uploaded, never sent to a server, never shared. There's no account, no login, no telemetry. It's strictly read-only — it never modifies your files — and secrets (API keys, tokens) are redacted the instant they're parsed, before they reach the index, the screen, or any export. The only optional network feature, an AI spec-check, is off by default and uses your own API key.
Who it's for
Indie developers and studios who live in Claude Code and want to stop bleeding tokens to loops they never saw. Engineers who believe verification — not generation — is the real bottleneck. Anyone who wants their long agent sessions to be legible instead of a black box.
The clarity of reading every transcript by hand, at the speed of a local analyzer — privately, on your own Mac.
Coming soon for macOS. One-time purchase, no subscription. macOS 13 (Ventura) or later, Apple Silicon or Intel.
© 2026 NYRAI LLC
© 2026 NYRAI LLC