Find the tokens your agent wastes. Prove the fix didn't hurt.
Only fixes that prove themselves on your own sessions get turned on. Open source. Runs on your machine.
brew install usetrove/trove/troveWorks with Claude Code. Sends nothing by default.
$ troveTrove report: this month (since Oct 1)Headline Spend ~$73.50 (estimated at list prices, not an invoice); 1.6M tokens (measured) Waste found at least ~$5.85, 8% of spend (estimated; overlapping findings counted once) Savings none yet; no optimizer has been on (estimated)Biggest leaks (estimated, overlapping, not additive) ← every number is labeled Repeated tool results ~$4.50 60K tokens Noisy tool output ~$3.00 90K tokens$ trove autopilot on - It starts in shadow mode: nothing is changed until there is enough evidence. - Only lossless optimizers that leave your cached prompt alone are eligible. - Changes reach only sessions that start afterwards, never one in progress. - Anything you set yourself in config.toml is left exactly as you set it. - A guardrail turns an optimizer back to shadow if it stops paying off. - It never touches your login or your traffic beyond what you already set up. - Nothing leaves your machine.[y/N] yAutopilot is on. It starts in shadow mode and changes nothing yet.$ trove autopilotduplicate_result_collapser shadow (collecting shadow estimates)noise_normalizer shadow (collecting shadow estimates)$ trove autopilotduplicate_result_collapser shadow (collecting shadow estimates)noise_normalizer on (promoted from shadow evidence, estimated)Changes reach only sessions that start afterwards, never one in progress.$ trove autopilotduplicate_result_collapser shadow (collecting shadow estimates)noise_normalizer shadow (guardrail: net saving went negative)Run trove to see where your tokens go. Example figures, not a real user.
One question. Nothing changes until there is evidence.
Each fix is tried on a copy first and records what it would have saved, after its own overhead.
Switched on only when the test shows a saving, only between sessions, and never over a setting you made.
If it stops saving, results get worse or errors rise, it goes back to testing and tells you. trove undo reverses everything.
Every number is labeled.
We ran 30 long Claude Code sessions with and without Trove and compared the cost.
Cost change with model routing
−25%estimated
Cost per session went down by a quarter, and every task still passed. Needs an API key.
Tasks passed
30 of 30measured
Every task passed, with and without Trove.
One pre-registered study on Sonnet. Model routing is optional: it hands tool-calling turns to a cheaper model. See the full results
Nothing changes until it's proven
- 1
Watch
Trove reads the session history your agent already keeps. It changes nothing.
- 2
Shadow
Each fix is tried on a copy of your requests, never the real ones, and records what it would have saved after its own overhead.
- 3
Promote
A fix is switched on only if that test shows a real saving, only for new sessions, and never over a setting you made.
- 4
Measure
Once on, it keeps being checked. If it stops saving or results get worse, it switches itself off and tells you.
Sends nothing anywhere by default.
Data leaves your machine only if you opt in with trove contribute enable. It never touches your login.
One command undoes it.
trove undo reverses everything Trove changed.