Katexs
Introduction

Get started

Introduction

Katexs is the infrastructure behind autonomous intelligence. Build voice and chat agents that answer calls, talk to website visitors, use your tools and remember every customer, on one tenant-isolated platform.

What is Katexs?

Katexs is a platform for building, deploying and operating AI agents that talk to your customers. A Katexs agent can answer and place phone calls, hold a conversation in the browser, chat on your website, look things up in your knowledge, book appointments, call your own APIs and hand off to a human, all while Katexs handles the realtime audio, the model calls, billing, security and logging.

You configure an agent once in the dashboard. Katexs runs it on every channel you connect, keeps one memory of each customer across voice and chat, and records every call, message and tool run so you can see exactly what happened.

Katexs platform architecture Channels connect through the Katexs edge to the voice and chat runtimes, which call tools in your systems. All runtimes share one tenant-isolated data layer. Channels Katexs edge Runtimes Your systems PhoneInbound and outbound Web callBrowser, short-lived token Website widgetwidget.js, streaming chat Your backendHTTPS API (preview) Edge & governance Auth & tenant check Origin allow-list Credit guard Rate limits Webhook verification Idempotency & dedupe Voice runtimeRealtime speech pipeline STT LLM TTS Carrier-grade telephony layer Chat runtimeServer-sent events, token by token Direct LLM Flow engine Knowledge retrieval per message Custom functionsYour HTTPS endpoints CalendarGoogle Calendar booking Human transferForward to your staff CRM & automationsIntegrations, webhooks Shared platform, isolated per workspace DatabaseRow-level security Credit ledgerAtomic, idempotent Shared memoryContacts, timeline Traces & logsEvery call and tool
How a request moves through Katexs: channels, the edge, the runtimes, your systems, and the shared tenant-isolated platform.

How a voice agent works

Every voice agent is a realtime pipeline of three parts, each of which you can choose:

  • Transcriber — turns the caller's speech into text (for example Deepgram Nova).
  • Model — the language model that decides what to say and when to use a tool (for example GPT-4o or Claude).
  • Voice — turns the reply back into natural speech (for example ElevenLabs or Cartesia).

Katexs streams all three at once so the agent can listen, think and speak with natural turn-taking, and lets the caller interrupt it like a person would.

Voice pipeline and latency targets Caller speech is transcribed, reasoned over by the model and synthesized back to audio. A tool call branches off the model and must answer within 800 milliseconds or acknowledge asynchronously. Callerspeaks TranscriberSpeech to texte.g. Deepgram ModelPrompt, memory, toolse.g. GPT-4o, Claude VoiceText to speeche.g. Cartesia Callerhears Tool callAnswer in <800 ms or ack async Your endpointPOST tool-calls Latency targets per turn <200 ms <300 ms to first token <150 ms to first audio
The voice pipeline for one conversational turn, with the latency target for each stage. Targets, not guarantees: provider and network conditions vary.

Building blocks

Voice agents

One system prompt, a model, a voice, a transcriber and tools. Best for most phone and browser use cases.

Chat agents

The same brain on your website, streamed token by token through the Katexs widget or the chat API.

Tools

Built-in calendar booking, call transfer and your own HTTPS functions, with strict latency budgets.

Knowledge & memory

Documents your agents answer from, plus a shared contact timeline across every channel.

Flows

Visual, multi-step conversation logic for chat agents that need structured paths.

Crews

Several specialist agents on one call, handing off with full context. Coming soon.

Choose your path

What makes Katexs different

  • Tenant isolation by design — every agent, number, call, contact and credit belongs to one workspace, enforced in the database and checked again on every voice event.
  • Atomic billing — usage is charged once per call or message on an idempotent ledger, so retries never double-charge.
  • Latency budgets — tools that block a live caller must answer within 800 ms or acknowledge asynchronously; the caller never hears dead air.
  • Cross-channel memory — a caller who chatted on your website yesterday is recognized when they call today, if you turn shared memory on.
  • Full traceability — every call, message and tool run is logged with timing and cost.

Common use cases

Use caseTypical setup
Front desk and receptionistVoice agent on your main number, calendar tools, transfer to staff
Appointment bookingVoice or chat agent with the built-in calendar tools
Lead qualificationVoice agent with call analysis fields and a CRM function
Website supportChat agent with a knowledge base and email handoff
After-hours answeringVoice agent with voicemail detection and call summaries

Tip

New to voice AI? Start with Quickstart: phone call, then read the Prompting guide. Most quality problems are prompt problems.

Last updated November 2026 · Found a problem? Email support@katexs.com