Otto is the lead. It takes any request, does the work or routes it to the specialist that owns it, and reports back. Specialists are yours to build: Otto interviews you, writes the persona, you approve it.
Finance gets its own persona, skills, toolbox, memory and home base.
“Build me a Finance agent.” Otto asks what lands on its desk, how it should sound, what it must never do, and when to pull you in. At most a handful of questions. You answer in plain words.
Otto drafts the agent: who it is, its mission, what it handles, how it talks, its domain rules, what it never does, when it escalates. Read it. Change a line. Approve. Every version is kept and any one can be restored.
The agent asks for the apps it needs. A connect card appears in the conversation. Click, sign in, and the agent carries on exactly where it stopped. Any app with an API key works too: paste the key, Otto finds the docs, works out where the key goes, tests it, and fixes it if the test fails.
“Every weekday at 8, chase every invoice over 30 days and draft the emails.” That sentence is the whole setup. No flowchart, no wizard, no template to fill. The drafts land as one card for you to approve.
Your agent now runs on a schedule, on incoming mail, on a change it is watching, on a goal you hand it, or when another agent passes it work. Ask “what did you do while I was out?” and get a briefing.
# triggered by schedule, weekdays 08:00, your time zone ✓ accounting.list_invoices overdue > 30 days 14 found ✓ business_profile.read voice, signature ✓ email.draft ×14 in your voice ✓ guardrails.check do-not-contact, cool-down 2 skipped ⏸ approval.request one card, 12 drafts # nothing sends until you say yes # cost of this run $0.031
Otto does the work or routes it to the specialist that owns it, and reports back. It acts first and asks at most one clarifying question.
It tells you what failed, what it could not verify, and what is a draft versus done. Ask what it can do and it answers for your setup: what is connected, what is not, which agents exist.
Rules every agent shares: brand voice, approvals and truthfulness apply to the whole team.
A skill is a playbook: when to use it and how to do the job. An agent loads one only when the moment calls for it, so it stays sharp instead of stuffed. Start from the library, or write your own in plain words.
# when to use it An invoice is more than 30 days overdue. # how to do the job 1 Pull the invoice and the last contact. 2 Draft a reminder in the house voice. 3 Batch every draft into one approval card. # never × Send without approval. × Chase anyone on the do-not-contact list.
Switch abilities on or off per agent. Give it every connected app or a hand-picked list.
Your agents read and send through your actual mailboxes on Gmail, Outlook or any provider, and through LinkedIn, WhatsApp, Instagram and Messenger. They watch those channels for inbound too. Not a widget. Your accounts.
The agent calls you when something urgent lands. With your approval it calls other people with a voice agent. Summaries and full transcripts.
CRM, email, calendars, project management, support desks, accounting, payments, e-commerce, documents, analytics, developer tools. Each connected app becomes a set of actions.
Add any remote server of your own, with no sign-in, a token or OAuth, and choose which of its tools ask first.
PDF, Word, Excel and CSV in your branding from your templates. Images and short video on brand. Video editing from clips you attach or generate.
Search and read pages, citing what was used. A real browser, live in front of you, for forms, logins and anything that needs a human hand.
No card. Your keys or ours. Nothing sends without your yes.