Connect your accounting and payments apps and the agent reads them directly. Every email it drafts waits for your yes. Every number it reports comes with where it came from.
The agent finds people and companies from a sentence, enriches and verifies them, checks what you already have saved before it buys anything, and writes every message in your voice. A batch of twenty is one approval.
Sends go out at a human pace, never twice to the same person inside the cool-down, never to anyone on your do-not-contact list, and a reply cancels every queued send to that person automatically.
Upload your FAQ, pricing and product docs once. Every reply cites those passages and the agent is forbidden from inventing product facts, prices or promises.
Anything that needs you personally gets a short holding reply and lands on your phone. Auto-replies are limited to the thread that wrote in.
Reports come out as real files in your branding. Watchdog jobs report only when something is wrong, so a quiet hour costs nothing.
LinkedIn outreach is paced and capped by the system, so the account stays in good standing.
Images come back as a grid with your brand palette already applied; you pick one, the agent drafts the caption, and publishing waits for your approval.
Video can be generated or edited: trim, stitch, resize per platform, speed, captions, thumbnails, GIFs.
“Watch my Gmail. Between 10 pm and 7 am, answer anything you can from my docs and escalate the rest.”
“Text me about anything from a customer. If an investor emails, call me, even at night.”
Importance rules in plain language. Incoming mail is screened against them cheaply. A match texts you, calls you, or both. Mark a rule urgent and it overrides your quiet hours; a non-urgent match inside them falls back to a text.
“Every weekday at 8, tell me what came in overnight, what is waiting on me, and what the team did.”
Posted to the home base and sent to your phone. Or switch on the daily review: what happened, what needs you today, one to three concrete suggestions for the business, and what was learned, written into the business profile every agent reads.
“Tell me when my competitor's price drops.”
“Track this account's followers every day.”
“Let me know when this product goes out of stock.”
A monitor is code, not a conversation. It fetches a page, a feed, a connected app or a scraper on a schedule, pulls out a few values and compares them with last time. Nothing new costs nothing, so it can run every few minutes.
Rules: any change, a number crossing a line, a move of more than N percent, a list gaining or losing items. When a rule fires, the change and only the change wakes the agent that owns it, or goes straight to your phone with no model involved. The agent builds a monitor by looking first: it probes the source, sees what comes back, and saves the monitor only once it has seen real values.
“Get quotes from five freight forwarders for this shipment and put them in a table.”
“Find us a venue for forty people on the 14th within budget and get a hold on the date.”
A mission is a goal that runs until it is done, stuck, or you stop it. The agent works it across as many runs as it takes, carrying its progress forward. At a login, a one-time code, a captcha or a payment, it hands you the page and resumes after. If it stops making progress it tells you instead of looping.
You watch it live: what it is doing right now, every run, every tool call, and the full audit log. Updates can go to your WhatsApp.
The agent opens pages, clicks, fills forms, scrolls, reads and takes screenshots. You watch live in the chat or in a docked pane. Take control at any moment for a login or a code, then hand the page back.
Risky clicks go through approval and the system makes the click, not the model. Agents never type passwords, codes or card details. Text on a web page is treated as data, never as instructions.
“Write the proposal for Northwind: three-month pilot, eighteen thousand, starting March.”
Save your best proposal as a template once. The agent follows its section order, writes real scope, deliverables, timeline and pricing, and the file comes out with your logo, colours, fonts, footer and legal line. It is forbidden from shipping a document with a placeholder left in; if a fact is missing, it asks. PDF for anything client-facing, Word when they will edit it, Excel with a frozen header and filters for data.
Anything you would type in the web app you can send to your Telegram bot or WhatsApp. Approve, reject, ask for changes with a note, check a batch's status, pause, resume, stop, or drop items, all from the chat you already have open. First answer wins across every channel.
One workspace per client, brand or business unit. Each has its own agents, data, connected accounts, keys and credits. Nothing is shared. The activity log becomes client-facing proof of work. Invite teammates with email-locked links that expire and can be revoked.
“What should I use for invoicing?”
Otto searches the app directory, the public server registry and community servers, recommends one, and offers to connect it from the chat. Not in the directory? Name the app. Otto finds the API's base URL and documentation, tells you where to get your key, wires it, tests it, and reads the error if the test fails. Give it the app's OpenAPI document and every endpoint becomes a tool with exact parameters.
Every connection also gets a webhook URL and an instruction in plain words: “when a payment lands, save the customer as a lead and text me.” Deliveries are checked against the app's signature, and an app with no webhooks gets polled instead.
Click any one to copy it, then paste it into Otto.
No card. Your keys or ours. Nothing sends without your yes.