Culpa - Local-first cost tracking for AI products
by•
Your bill spiked and your dashboard shows one number. Culpa names the conversation that spent it, then forecasts your next bill. Local-first LLM cost tracking. Your prompts stay on your own infrastructure.
Most AI dashboards just show what you spent. Culpa shows who spent it. As a local-first cost tracker, Culpa traces every LLM dollar back to the exact conversation and user. The best part? Your prompts never leave your infrastructure.
Replies
Hey Product Hunt,
I’m Zolani, founder of Culpa. Culpa started with a question I couldn’t get answered while building a different AI product:
What is this bill increase? What exactly moved it?
Provider dashboards showed the model, tokens and total spend. That was useful, but when we started looking at unit economics, I wanted to know which conversation, feature, customer or agent step was behind the increase. I kept going back to spreadsheets, logs and guesswork, so eventually I built the tool I wanted for myself.
One conversation made 2,438 model calls.
🔍 Our launch trace recorded 6,621 model calls. One conversation accounted for 2,438 of them. Culpa calculated $946.43 in model cost from the recorded usage and applicable provider pricing.
Then we followed the conversation.
Turn 1 carried 30,157 tokens of context. Turn 2 carried 62,411.
By turn 18, it had reached 164,926.
The final context reached 867,161 tokens.
By turn 2, the context being carried forward had already roughly doubled. The conversation was still running comfortably inside the model’s context window, but its cost profile had changed very early.
💡 Culpa lets you work backwards from there.
Which customer used which feature? Which conversation ran? Which workflow expanded? Which call or step changed the pattern? Instead of knowing only that your model bill increased, you can see which part of the product needs attention.
You can also compare AI cost with customer revenue and forecast what a feature, cohort or launch is likely to cost before you run it.
🔓 Culpa runs on your infrastructure. Your prompts and responses stay there. Culpa only sends the counts needed to run your plan, and if Culpa ever fails to record something, your application keeps running.
For the PH launch, we’re opening the first 100 Culpa workspaces at $199 for the founding year, normally $480.
That’s seven months free. Find your first culprit and activate within 14 days to keep the founding price. After year one, your workspace renews at the standard Builder price.
First 100 workspaces only.
One question for the thread:
When your AI workflows get expensive, what’s the smallest unit you can currently trace that cost back to?
Model, request, conversation, customer action or individual call?