What Katara does not do

Where your money is

Where the money sits depends on how you pay.

Account credits are a balance Katara keeps for your account. A request paid with credits holds what it might cost while it runs, then is charged for what was delivered, and the rest is released; a request that fails is charged nothing. Credits pay for Katara’s own models and, while Katara Community is on, for Community providers, which Katara pays itself so the provider never learns who you are. Purchased credits do not expire. Signup credit is spent first, expires 90 days after it is granted, and cannot be withdrawn or refunded.

USDC stays in an account only your passkey controls, and the prepaid balance in the settlement contract can only be debited for replies you received or paid back to you: the contract has no function that sends it anywhere else. The contract can be upgraded by its owner, Katara, and the upgrade is a public transaction on chain. Withdrawals pay out after a one-hour delay so replies in flight are charged first; nothing is ever served against money that is not already in the contract.

What the system will not do

Two rules hold whatever you pay with. They are enforced in code and covered by tests that must keep passing.

Your prompts and replies never go on-chain or into our logs. Some Katara features keep them: the chat app keeps your conversations under your account, and any other feature that needs content keeps it in a store of its own, with its own retention period. The marketplace itself records tokens, price, timing, and who served the request. Providers on Katara Community see the request they serve and nothing about who sent it, and they may not keep it.

Providers never learn who you are. Requests reach a provider under a job ticket, not an account. The provider is paid from the settlement contract, not by you.

What Cortex reads

When Katara Cortex serves a request, the text of its latest message, cut to its first 300 words, goes to a decision model that Cloudflare runs for us on Workers AI. It answers two questions: does the request need step-by-step reasoning, and should a light, standard or deep model answer it (Light, standard and deep). The answers decide whether the model thinks before replying and which model fits. In Katara’s chat apps it is also asked whether the message is about Katara, which adds passages from Katara’s documentation, and whether it asks which model or company is behind the assistant. Cloudflare already carries every request on its way to Katara, and it does not store or train on what the model reads. A request that sets thinking itself skips the first question, and one with a video skips the second, since a video always goes to the deep tier.

What is verified

Every reply is backed by a receipt the provider signs. The router checks the receipt against what it observed, token by token, before anything is settled. A receipt that does not match is not paid, and the provider is struck.

Security work

The contracts and services were reviewed against adversarial scenarios before launch, with every high-severity finding fixed. Report a vulnerability to security@katara.com.