# Katara

> OpenAI-compatible inference from Katara Cortex, paid per reply with account credits or USDC.

Web page: https://docs.katara.com/

Chat and build on Katara Cortex. Pay per reply with account credits, or in USDC from a wallet you control.

- [Start chatting](https://docs.katara.com/start/chat.txt)
- [Use the API](https://docs.katara.com/start/api.txt)
- [Build agents](https://docs.katara.com/products/overview.txt)
- [Provide inference](https://docs.katara.com/providers/overview.txt)

**One price per reply**

Every reply shows what it cost. There is no subscription and no minimum. New accounts start with a signup credit.

**Credits or USDC**

Account credits are bought by card. USDC stays in a wallet only you can spend from: approve a cap once and each reply pulls its exact cost. Replies paid either way can run on Katara Community. [Pricing and fees](https://docs.katara.com/concepts/pricing.txt)

**Verified models**

A model on Katara is a pinned bundle: the same weights and settings on every provider, checked by validators that probe providers with known prompts.

**Always an answer**

If no provider can take a request, a managed fallback answers at a published price. The reply says so.

## Three products

Katara is three products on one account: the same organizations, environments, keys and API. Each stands on its own, and none of them requires the others.

**Cortex: inference**

Models behind one OpenAI-compatible API. Requests paid with account credits or USDC run on Katara's own models and can also run on independent providers in Katara Community. Cortex answers requests; it keeps no agent state. [How Cortex answers](https://docs.katara.com/concepts/cortex.txt)

**Synapse: agents that run and remember**

Durable execution for agents: runs that can be inspected, cancelled and scheduled, knowledge and memory an agent draws on, and evaluation that measures whether it is doing its job. [Synapse and Locus](https://docs.katara.com/products/overview.txt)

**Locus: what an agent may do**

Authority over actions: the grants and budgets an agent holds, the tools it can call, the policy that decides each action, the approvals a person gives and the governed action that runs after them. [Tools and approvals](https://docs.katara.com/products/tools-and-approvals.txt)

They meet where one needs what another does. An application can call Cortex with no agent involved. A Synapse agent asks Cortex for inference, and when it reaches outside itself, by calling a tool or spending a budget, Locus decides whether it may. An agent you run yourself can act through Locus without Synapse at all.

Every account has a Test environment and a Live one. [Networks](https://docs.katara.com/about/networks.txt) says what each settles on, and [Product availability](https://docs.katara.com/products/availability.txt) how to ask whether a Synapse or Locus service is switched on where you call it.