Concepts
How Katexs works
The architecture behind Katexs, what happens on every call and message, and the guarantees the platform makes.
The layers
Katexs sits between your customers and the AI. Requests enter through a channel (a phone number, a browser session, the website widget, or in future your own backend). Every request passes the edge, where Katexs authenticates it, finds the workspace it belongs to, checks credits and rate limits, and verifies signatures. It is then served by a runtime: the voice runtime for realtime speech, or the chat runtime for streamed text. Runtimes call your systems through tools, and read and write the shared platform: the database, the credit ledger, shared memory and traces.
Workspaces
A workspace is the unit of isolation. Agents, phone numbers, calls, chats, contacts, knowledge, tools and credits all belong to exactly one workspace. Team members are invited into a workspace with a role (owner, admin or member). Nothing is ever shared between workspaces. See Tenant isolation.
What happens on a phone call
- 1
The call connects
If the number has an agent assigned, the voice engine starts it. If not, Katexs is asked which agent should answer and must reply within 7.5 seconds. Katexs aims for under 800 ms.
- 2
The call is logged
On the first status update Katexs creates the call log under the right workspace.
- 3
The conversation runs
The agent listens, thinks and speaks in realtime. When it needs a tool, the tool request is executed and the result is spoken back.
- 4
The call ends
The end-of-call report arrives with the duration, transcript, recording link, summary and analysis.
- 5
Katexs settles it
The call is charged once on the ledger, the transcript is stored, the caller's contact and timeline are updated, notifications fire and a trace is written.
What happens on a chat message
The widget loads its configuration, then posts each message. Katexs checks the origin, the signed visitor token, rate limits and credits, reserves the cost, and streams the model's reply back as server-sent events. If the reply completes, the usage is committed and the transcript saved; if it fails or the visitor disconnects, the reservation is refunded.
Guarantees
| Guarantee | How it's enforced |
|---|---|
| No cross-tenant access | Row-level security on every table; workspace tags on every voice resource, checked on every event |
| No double charges | One ledger entry per call id or message id; duplicate events are detected and dropped |
| No silent failures | Every failed tool returns a calm, speakable message and is logged with its reason |
| Secrets stay on the server | The browser only receives short-lived, single-purpose tokens |
| Every action is traceable | Calls, messages and tool runs write traces with timing and credit cost |
Last updated November 2026 · Found a problem? Email support@katexs.com
