Cut your LLM bill.
Keep the speed.

See what you can recover

Point it at a coding-agent session and it answers in dollars — recoverable, never "wasted":

That number is the point. Most sessions have one — the question is how big, and whether you act on it.

How it reduces LLM costs

It detects the common, expensive token-waste patterns on-device — each with a copy-paste fix:

Redundant round-trips

The agent re-polls a tool whose result didn't change.

Stuck loops

Repeated output that isn't making progress.

Oversized context

Large prompts paying full price every turn instead of cached.

Uncached repeated context

The same big context re-sent, uncached, call after call.

Missing env bootstrap

Early turns burned re-discovering the environment.

Local-first and values-free: it works on token counts, costs, and equality-only hashes. Your prompts and code never leave the machine.

Two halves of one bill

Cost is what your AI spends. Your data is what it can cost you. Latten watches both: @latten/cost finds the spend you can recover, and @latten/redact detects PII and secrets leaking into your LLM calls — on-device, types and counts, never values. Same local-first engine, pointed at the other risk.

See both packages → · How your data stays safe →

Reduce costs in your production app

Pricing

Flat and predictable. We never charge by the call — metering your spend would mean charging you more to be more efficient, which is backwards. Your first week is everything, free. After that, knowing where you stand stays free; understanding and acting is paid.

Enterprise for larger orgs — talk to us.

Honest about what it is

Detection isn't magic — we catch the patterns we know, and we label fidelity plainly (measured vs estimated). The savings ledger only ever counts fixes you actually applied. We'd rather under-claim and keep your trust than inflate a number.