Bring your own AI

Your model, your key

Point Kalagh at the provider you already use — or at a model running on your own hardware. Nothing here is wired to a single vendor, and your customers’ conversations are never used to train anyone’s model.

Twelve kinds of provider

Hosted or self-hosted, every feature works the same way — because the platform talks to a model, not to a company.

OpenAI
Anthropic
Google Gemini
Azure OpenAI
AWS Bedrock
Mistral
Groq
DeepSeek
xAI Grok
OpenRouter
Any OpenAI-compatible endpoint
Self-hosted Ollama or vLLM
AI

AI where it helps, off where it does not

You decide which provider answers, what it is allowed to do, and how much it may spend doing it.

Your key, under your control

Paste a key, test it, list the models it can reach and choose a default. Keys are stored encrypted.

  • Test the key and list the models it can reach
  • Keys encrypted at rest
  • A default model per organisation

A budget you set

A monthly token budget and a per-conversation cap, with a usage view so nobody has to guess what it cost.

  • A monthly token budget
  • A per-conversation cap
  • Usage you can actually look at

An AI agent inside a playbook

Drop an agent into a conversational flow and choose exactly which tools it is allowed to use.

  • Tools you pick: tags, contact fields and variables
  • Search messages, read the contact, call a sub-playbook
  • A simpler AI auto-reply, when a whole agent is too much

A human is always reachable

A keyword is matched before the model ever sees the message, and the model can ask for a human itself. Either way the team is told.

  • Keyword handoff, checked before the model runs
  • The model can hand off on its own
  • A fallback path when a provider is down or the budget is spent
  • A full trace: the prompt, the answer and the tokens

Privacy, stated plainly

The three things people ask before they turn any of this on.

Never used for training

Your customers’ conversations are not used to train models — not ours, not anybody else’s.

Self-hosting is a first-class path

An open-source model on your own hardware is a supported choice, not a workaround for one.

Every run can be explained

The prompt, the answer and the token count are kept with the run, so an odd reply is something you can look into.

Add AI on your own terms

Bring a key or a self-hosted model, set a budget, and keep the handoff to a human.