Cloak – let AI agents use your API keys without ever seeing them
No read_secret tool means agents can use keys without ever reading them.

Read-only admin keys beat proxying traffic for cost tracking without the latency risk.
Indie hackers and developers running multiple AI apps
Helicone · LangFuse · Portkey
Hi there, over the past 1-2 years I've built and experimented with many AI apps. Some are in production, some died in the experimental stage.
The ones in production currently costs me ~$500 a month to run – using a variety of models from OpenAI's Whisper to Anthropic LLM.
Because it's important for me to track the Cost of Goods Sold of each individual app, the strategy I've used so far is to use a different credit card for my few apps. However, that obviously does not work the more apps I build.
This is why I built Costbase – a simple app that just requires that I paste in an admin key with no code change (and no proxy sitting between my app and the AI provider) – and show me the cost of my AI bill broken down by apps and not API keys.
I intentionally kept this product as simple as possible – something I would actually like to use. I know there are observability tools like Helicone that are much more capable. But to be honest I didn't even try out the competitor and just built something I'd like to use myself.
I'd like to hear from the community to see if anyone feel a product like Costbase is genuinely useful.
Let me know your thoughts in the comments.
(I'm in Taipei so going to sleep in about two hours. Will be back tomorrow morning to answer any additional comments)
No read_secret tool means agents can use keys without ever reading them.
Users bring their own API keys so indie devs skip inference bills entirely.
Agents never see real keys, but Vault already does secret injection.
Proxy tokens worthless if leaked, real keys never enter LLM context windows.
Drop-in Claude API proxy with real-time cost dashboard—but Anthropic's own billing UI exists.
Not a proxy — optimizes Anthropic and OpenAI caching without adding latency or seeing your data.