Built like a colleague, not a chatbot
Every part of Kavvo exists to make one thing true: you talk, work gets done, and you can see exactly how. Here is the machinery, piece by piece.
Chat-first, with a living orb
Kavvo has one main room: the conversation. At its centre sits the orb — a living indicator that breathes when idle, leans in when listening, churns while thinking, and pulses as it speaks. Long jobs don't hide behind a spinner; the orb and the session meter keep the state of the work visible at a glance.
The header always shows the truth of the session: which provider and model are connected, how much context is in play, and what the conversation has cost so far. No silent model swaps, no mystery bills.
A real listen → think → speak loop
Voice in Kavvo is not a dictation box bolted onto chat. Tap the orb and talk: Kavvo listens until you finish, thinks with its state visible, then speaks its answer aloud. Interrupt mid-sentence and it stops talking and goes back to listening — the way a person across the desk would.
Because voice and chat are the same conversation, anything said aloud lands in the same thread, memory, and activity stream as anything typed.
Your provider, your model, your prices
Kavvo is BYOK — bring your own API keys. Connect a provider and Kavvo fetches that provider's real model list on the spot; there is no stale built-in catalogue. Pin a specific model when the work demands it, or choose Auto and let each task be routed to a sensible model.
Keys stay on your machine or server and call the provider directly. Kavvo never proxies your requests, never sees your bill, and adds no markup — your provider statement is the only statement.
A real-human browser, driven for you
Research, price checks, form fills, multi-page lookups — Kavvo operates a genuine browser, the same kind you use, rather than scraping raw HTML. Pages render, scripts run, and results come back as findings in the chat, with the trail visible in the activity stream.
When a task needs your logins, Kavvo can attach to your own browser session instead, working inside the accounts you already have open — under your watch, in the stream.
Long-term memory, governed by your constitution
Kavvo remembers across sessions: your projects, people, preferences, and the facts that make its work specific to you. That memory is not a black box — it is guided by a constitution, a short document you own that sets what Kavvo should remember, how it should behave, and what it should never do.
Edit the constitution any time and the behaviour follows. Ask what Kavvo remembers about a topic and it will show you, plainly.
Skills it creates for itself
The defining idea of Kavvo: capability is not a catalogue you shop in. When a task needs something new — a data pull, a document routine, a multi-step errand — Kavvo designs and writes that skill itself, tests it in the background, and keeps it for next time.
Skills accumulate into a personal toolkit shaped by your actual work. They are never a menu you manage, a store you browse, or a subscription you renew.
Briefings before you ask
Recurring work becomes quiet routine: a morning briefing that gathers what matters, a weekly review that shows what got done and what's drifting, overnight delegations Kavvo works through and reports on by morning.
Every automated run ends with the same honesty as a live chat — a summary of what happened, what it touched, and what still needs you.
Forgetting, with a receipt
Memory is only trustworthy if leaving it is easy. The privacy guardian can purge an entire topic from Kavvo's memory — every fact, note, and association — and issues a receipt listing what was removed so you can verify the forgetting happened.
It also watches the exits: before anything is shared outside your install, content is scanned for secrets like keys and passwords, and flagged rather than sent.
See it running, not just described
The live preview instance is open. Connect a key and put every feature on this page to work.