Request access

Make every token count.

Stop paying for the tokens your coding agent wastes.

See what your sessions cost How it works
acme/api/Sessions/ses_8K2QF7

Migrate billing to the new payments API

Claude Code3 h 12 min412 tool calls

You paid
$38.40
Pending verification
−$21.05
Verified saving
−$5.25
What Trove would change
  • retry-stop

    pytest failed the same way 14 times in a row. Stop after 2 and show the agent the diff instead.

    −$11.20 Ready to verify
  • dedupe

    Repeated test logs. 4.1 MB kept out of context, all of it recoverable.

    −$5.25 Verified
  • cache

    Prompt edits broke the cache 4 times. Pin the system prompt for the rest of the session.API / Bedrock / Google Cloud

    −$9.85 Ready to verify

Example session. −$5.25 is verified. −$21.05 is still waiting on verification.

  • Local-firstRuns on your machine
  • ReversibleAnything Trove cuts is saved and can be restored
  • Usage billingAPI keys, Bedrock, or Google Cloud

Long sessions get expensive quietly.

Coding agents carry the whole conversation forward on every step, so repeated and bulky output keeps adding to the bill, even with caching.

Trove trims the waste and keeps everything it trims recoverable on your machine.

Sometimes a fresh session is the right call. When it isn’t, Trove helps keep the long one from getting wasteful.

Doesn’t prompt caching fix this?

Request size, turn by turn (illustrative)
  1. turn 482 KB
  2. turn 8145 KB
  3. turn 12207 KB
  4. turn 19326 KB
  5. turn 27451 KB
  6. turn 36582 KB
Rest of the conversation Repeated output

Illustrative example: by turn 36 the request is 238 KB instead of 582 KB.

Nothing changes until it’s proven.

Every fix is tested on your own tasks before it’s turned on.

  1. 01

    Watch

    Tracks what each session costs. Changes nothing.

  2. 02

    Suggest

    Spots waste, like repeated logs or retry loops.

  3. 03

    Verify

    Reruns the tests from that session with the fix in place. Those tests have to succeed, or the fix stays off.

  4. 04

    Turn on

    Fixes that pass start saving. Retested after updates.

trove verify --policy reread

Verify reruns that session’s tests with the fix on, then compares the run to the one without it. The fix turns on only if the tests still pass and the bill goes down.

  • Task successSameThe tests still pass.
  • Tokens sent−22%The prompt got smaller.
  • Cost per verified task+5%The bill went up, so the fix stays off.

RejectedFewer tokens sent. More dollars spent.

After a claude-code upgrade, that result is stale. The check blocks the merge until verify runs again.

We optimize one number: billed cost per verified successful task.

A smaller prompt can cost more if it breaks the provider’s cache. Trove measures dollars, not tokens.

−22%Input tokens sent
+5%Cost per verified task

Example: the reread policy from the verify run above. Trove rejected it.

What Trove won’t do.

  • Rewrite context with an LLM

    Deterministic rules only. Same input, same output.

  • Resell tokens

    You keep your key, and your bill.

  • Touch your login

    Trove never sees or forwards your Claude credentials.

  • Change things silently

    You see what would change before anything turns on.

Frequently Asked Questions

Who is Trove for?

Developers and teams who pay per token for Claude Code. It saves the most on long, test-heavy sessions like test-fix loops, migrations and CI agents. Claude Code comes first. Cursor, Aider, OpenCode and Codex CLI are next.

What counts as a task?

A task is one goal you give your coding agent, from request to done, like a bug fix, a feature, or a migration. Trove can verify the ones that end in an automatic pass or fail, like a test or a CI job. Bug fixes and migrations usually have one. A feature does if it has tests. Other work is still tracked, but Trove can’t prove it succeeded.

Does it work on a Pro or Max plan?

No. Early access is for usage-based billing: an API key, Bedrock, Google Cloud, or Team and Enterprise usage credits. That is where a saving shows up on the bill. Pro and Max are subscriptions, so they are not who this is for. You still sign in to Claude Code as usual, and Trove never sees your login.

How does Trove plug in?

It runs inside Claude Code without modifying Claude Code. Usage data shows what each request cost. It trims repeated logs and stops retry loops. Everything stays on your machine, and admins can roll it out to a whole team through managed settings.

Won’t trimming break my agent?

It can, so Trove checks. Claude’s replies won’t be word-for-word the same, but Verify confirms your tasks still succeed, and fixes are retested after updates. Anything Trove cuts is saved on disk, costs no tokens to keep, and comes back if you turn a fix off. Verify does use tokens on your own key when it reruns tasks, once per fix and again after updates.

Doesn’t prompt caching already fix this?

It helps a lot, and Trove is designed to work alongside it. Cached tokens are discounted, not free, and caches expire or get rebuilt over a long session, which is where costs build up. Trove measures what you actually pay, so you can see whether a change lowers your bill and not just your token count.

Get early access. See what your sessions actually cost.

Free while we’re early. If you pay per token, we can look at your sessions with you.

I read every request and reply myself.