Train it on your documents, put it on every channel, and keep your team one click away for the conversations that need a person.
Answers grounded in your PDFs, Word docs and text files — with citations your team can audit.
The AI knows when to stop. Your team takes over on the customer’s own channel.
A configurable form the agent opens at the right moment. Deduped, exportable.
Speak to the widget; the agent talks back. Answer phone calls with a neural voice.
One agent on your site, Gmail, WhatsApp, SMS, Telegram, Slack, Discord, Messenger, phone.
Owner, admin and member roles, email invitations, and an escalations queue with assignment.
Everything the dashboard does has an endpoint. Keys are hashed, shown once, revocable.
Encrypted credentials, hashed passwords, signed webhooks, domain allow-lists, rate limits.
A monthly AI-spend ceiling per workspace with automatic downgrade to a cheaper model.
Upload PDFs, Word documents, and text files; the agent answers from them — with inline citations your team can audit in the dashboard, and an honest "that's not in my sources" instead of a guess when the answer genuinely isn't there. How grounding works →
When a customer asks for a person, when a question isn't in the knowledge base, or when your own trigger rules fire, the conversation escalates: the AI mutes on that thread, workspace owners and admins get an email, and the conversation waits in the escalations queue with priority, assignment, and private team notes. A teammate replies from the dashboard and the customer receives it live on their own channel — widget, email, SMS, WhatsApp, Telegram, Slack or Messenger. Resolve the thread and the AI takes the next conversation again. Business hours are respected: after hours, customers get your offline message instead of a promise nobody is there to keep.
A configurable form — name, email, phone, plus any fields you add — that appears when it should: before the chat for new visitors, from a header button, when the conversation shows buying intent, or when an escalation needs contact details. Repeat submissions merge instead of duplicating. Lead capture in detail →
Visitors can hold the mic button and talk to your widget; the agent transcribes, answers, and can speak the reply in one of six voices. On the phone, the agent answers inbound calls in a neural voice and every call becomes a text transcript — no audio recorded. Voice AI in detail →
The same agent — same knowledge, same persona — answers on your website, in Gmail, over SMS and WhatsApp, on Telegram, Slack, Discord and Messenger, on the phone, and through a signed webhook from your own systems. Tone adapts per channel automatically: email reads like email, texts stay short, Slack speaks Slack. See all integrations →
Every dashboard action has an endpoint — the embedded widget itself runs on the same public API. Keys look like cc_live_…, are stored only as hashes, shown once at creation, rate-limited per key, and revocable instantly. Read the docs →
Integration credentials are AES-256-GCM encrypted at rest, passwords are bcrypt-hashed, API keys are stored as SHA-256 hashes, payments run entirely on Stripe-hosted checkout (card numbers never touch our servers), and the widget respects per-agent domain allow-lists and rate limits. Full security page →
Each workspace has a monthly AI-spend ceiling. Every message logs its exact model cost; as spend approaches the ceiling, the router automatically downgrades requests to the cheapest model in the chain rather than surprising you with a bill. Message quotas per plan are enforced the same way — visibly, with headers, never silently.
The pieces are one workflow: your documents ground the answer, the citation proves where it came from, and returning-visitor recognition (the widget remembers who a visitor is across sessions, not what the model imagined about them) keeps the greeting personal. When grounding fails or a customer insists, escalation hands the thread to a person with the transcript attached. Everything else — channels, voice, leads, the API — is that same loop meeting your customers wherever they are.
OpenAI GPT-4o and GPT-4o mini on every plan; Anthropic’s Claude Sonnet 4.6 and Claude Haiku 4.5 on Pro. Model prices are tracked per message, and the cost ceiling downgrades to a cheaper model automatically as you approach it.
Answers are retrieved from your uploaded documents and cited inline; the agent is instructed to say plainly when the answer isn’t in its sources rather than invent one — and you can have exactly that situation trigger a human escalation.
No. Your documents and conversations are sent to the model providers for inference only — there is no fine-tuning or training of any kind in the platform.
Yes — the free plan lets you build agents, upload documents, and test in the playground and website widget. Cards come in only when you pick a paid plan, and the first charge lands after the trial.
10 minutes from sign-up to your first answered customer message. Free forever plan, no card required to build.