We build Kibble

Claude Code + Codex usage for small teams. We share weekly cost checks and collector notes. Open-source collector. Counts, never content.

[We build Kibble @use_kibble]

Kibble is command center for what every coding agent on your team spent this week: per engineer, team, model and repo
See the cost, catch the waste before the invoice does, and learn which workflows are actually paying off.
Our collector is OSS

[We build Kibble @use_kibble]

@FrankYouChill Advisor Mode saves on the easy half and pays twice on the hard half. A run that stalls has already burned the cheap model's tokens before the expensive one gets a look.

[We build Kibble @use_kibble]

@FelipeFr1702 Nobody there is defending the company, they're defending their own judgment. You can't tell a bad week of prompting from a quietly smaller allowance, so people pick the version that flatters them.

[We build Kibble @use_kibble]

@Blackwellboy @u1tra_instinct Breaking the context estimator is the interesting part. Every meter assumes context only grows, so the moment you free room mid-session the number people watch goes wrong.

[We build Kibble @use_kibble]

@imjustnewatai Service tiers are the quietest price lever there is. This one moves in a PR instead of a pricing page, and nothing in your usage tells you which tier a run actually got.

[We build Kibble @use_kibble]

@sakatayasha Trimming tool output and persisting memory across sessions pull opposite ways on the bill. One cuts what you send, the other guarantees you send it again tomorrow.

[We build Kibble @use_kibble]

@Ronny_MiniMax The budget block counts seconds and steps but not tokens, which is the leak you opened with.Compaction saves at the checkpoint and spends it again on re-reads, and neither of those shows up in wall time.

[We build Kibble @use_kibble]

@iamcheyan Two separate weekly pools means running out isn't one fact any more. You have to ask which pool before you know whether the week's actually over.

[We build Kibble @use_kibble]

@dorkitude Everyone guesses which part of their context costs the money, and the guess is almost always the prompt. Takes an audit to find out it was reading.

[We build Kibble @use_kibble]

@sashikantsingh_ Same tier is the trap in that comparison. One price on both doesn't mean the same amount of work, and you only find the gap by running dry on one of them.

[We build Kibble @use_kibble]

@Charlie_no_site The long project threads being the bulk of it is the real finding, and /clear is free to say and hard to do.People keep those threads because rebuilding the context is exactly what they're avoiding.

[We build Kibble @use_kibble]

@neuralkian Straight swap of a silent downgrade for a hard stop. At least you can see that one coming.

[We build Kibble @use_kibble]

@muscle_coding The expensive half of that workflow never shows up in the Codex limit. Writing the prompt on Fable or Opus moves the cost onto a different bill, it doesn't remove it.

[We build Kibble @use_kibble]

@Marktechpost Truncating tool results at 1,500 tokens is where a lot of that 77% comes from, and it's a trade, not a free win. Cheaper runs plus more retries can land in the same place.

[We build Kibble @use_kibble]

@soycronus Agree, and the awkward part is every vendor publishes a price per million and nobody publishes what their harness does with it. You find out after a month of running it.

[We build Kibble @use_kibble]

@YoussefHosni951 The part worth stealing is the hook, not the 90%. Routing rules in a config file are a suggestion, and the model treated them like one.

[We build Kibble @use_kibble]

@darrxgh Odds are the limit didn't move, your sessions got longer. Every extra turn re-sends the whole conversation, so the same task costs more late in a session than it did at the start.

[We build Kibble @use_kibble]

@chiraldevai Every price cut just raises the ceiling on what you'll let it try.

[We build Kibble @use_kibble]

@_nilni A bar only tells you how much is gone. The useful version says which model and which session took it, otherwise you're just watching a number drop.

[We build Kibble @use_kibble]

@hubeiqiao Weekly is the worst window they could have picked. You ration for four days then scramble to spend the rest, and the actual work doesn't arrive on that schedule.

[We build Kibble @use_kibble]

@davis7 The over-200K doubling is the part that'll get you, not the base rate. An agent crosses that line on its own in a long session and nothing tells you which turn did it.