Theo Agent gives you workers: colleagues made of AI that you describe in plain language and that keep working when you close the laptop. You tell a worker what its job is. It plans, acts, shows you each step beside the conversation, and asks when its approval policy requires it. Its tools and connections determine what it can do. Review important results and keep approvals on for consequential work.
Explore Theo Agent for a short overview. For explicit triggers and step-by-step automations, use Workflow Builder.
/theoclaw. Create one from /theoclaw/new. Each worker opens inside Theo Chat: pick it from the Theo Agent tab in the chat sidebar, talk to it, watch its work on the canvas beside the conversation, and change anything from the gear in its header.Web, Sidekick and your Mac
The web experience is in beta where enabled, with an eligible plan and AI credits.
The Theo Agents tab in Sidekick is coming soon. Preview cloud conversations, activity, questions and approvals from the desktop app. Advanced setup, connections and rich workspaces open on the web.
Creating or approving a cloud agent does not grant access to your Mac. Native Sidekick assistance uses separate permissions and shares your foreground mouse and keyboard.
Native Mac assistance is a separate beta. Screen access needs Screen Recording permission; clicking and typing need Accessibility. Downloading Sidekick does not unlock its unreleased Agents tab.
What can I ask it for?
The kinds of work it takes on. Each card names what the worker needs and gives you a sentence to type; the same sentences are one tap away in the setup wizard and in a new worker's chat.
Inbox and messages
Triage what came in, draft the replies, book the meetings, and hold anything that goes out until you approve it.
Needs: A connected mailbox, or the worker's own address.
- Run my day
“Look at my inbox and calendar for today, tell me the three things that need me, and draft what I should send.”
Start a worker for this - Morning inbox triage
“Go through my inbox from the last 24 hours. Tell me the three things that need a reply, and draft each reply for me to approve.”
Start a worker for this - Draft and send for me
“Draft a follow-up to the people who wrote in this week, show me each one, and send the ones I approve.”
Start a worker for this - Book my meetings
“Find a 30-minute slot next week that works for me and Sam, propose it to Sam by email, and book it once they say yes.”
Start a worker for this
Research and writing
Read the web, pull out what matters, and write it up as a note, a document or a brief with sources.
Needs: Nothing to connect. Add your company wiki for answers grounded in your own knowledge.
- Research and summarize
“Research the three strongest competitors in our space and write a one-page comparison as a note, with sources.”
Start a worker for this - A morning brief
“Give me a five-item brief on what happened in my market this week, one tight paragraph each, with a source for every item.”
Start a worker for this - A Monday recap
“Pull last week's numbers from my connected apps and write me a two-minute recap: what moved, what stalled, and the three things that need my attention this week.”
Start a worker for this - Write in my voice
“Turn these notes into two LinkedIn posts in my voice. Show me both before anything is posted.”
Start a worker for this
Watch and alert
Check a page or a number on a schedule and message you the moment it moves, without spending anything while it does not.
Needs: The page or dashboard to watch, and a channel to reach you on.
- Watch a page
“Watch https://example.com/pricing and tell me the moment anything on it changes.”
Start a worker for this - Ops watch
“Check our connected dashboards now, tell me which numbers are outside their normal range, and keep checking every hour.”
Start a worker for this
Build and ship
Build in a repository of its own: write the code, run the tests, keep the history. Publish it to GitHub when you want it there.
Needs: Nothing to start. A connected GitHub account only when you want the work published there, or to work in a repository you already have.
- Work on a repository
“Open owner/repo, fix the failing tests, and open a pull request with what you changed.”
Start a worker for this - Keep a repository healthy overnight
“Look at owner/repo and tell me the three most valuable things you could ship tonight inside this brief: thin test coverage, flaky tests, and open issues labelled for you.”
Start a worker for this
Its own computer
Turn one file into another, run a program, fill in a website, and leave the finished work in your Drive.
Needs: Nothing. Every worker can open its desk; sign-ins happen on the live page with you watching.
- Turn this into a spreadsheet
“Take the PDF I am attaching, pull the table out of it into a spreadsheet with a totals row, and leave it in my Drive.”
Start a worker for this
Create and publish
Turn your ideas, notes and recordings into posts, decks, images and short videos in your voice, and publish only what you approve.
Needs: Nothing to start. Connect the accounts you want it to post to.
- Show notes and quote cards
“Here is my latest interview transcript. Pick the three best moments, write the show notes, and make a quote card for each one for me to approve.”
Start a worker for this - A weekly content pack
“Turn these notes into a five-slide carousel, one striking image and a twenty-second video script in my voice.”
Start a worker for this
Your first ten minutes
The wizard at /theoclaw/new is five steps. Only the first one creates anything; every later step edits the same worker, so nothing you set is lost if you leave and come back.
Describe
Name it and say what you want done in plain language, or leave the job blank for a general assistant. Tap one of the example jobs if you would rather start from one. Draft with Theo writes the goal and persona for you.
Tools
Choose the tools this worker may use. Available actions depend on its tools, enabled capabilities and connected accounts. Its approval policy determines when it pauses to ask you.
Connections
Connect your accounts once and every worker can use them from that moment. This step is where you check what is connected and turn an account off for this worker if you would rather it did not use one. Channels such as Telegram, email and text live here too.
Brain
Let Theo pick the engine behind each step (the recommended start), go economy, or pin one engine. You can change this any time from the worker's composer.
Launch
Review the setup, decide when it asks and when it runs, then open it in Theo Chat. It speaks first: what it understood, three things it can do right now, and what to connect next.
Then just talk to it. Say what you need in your own words, think out loud, or change direction. It plans, acts, and shows each step beside the conversation. Type / in the composer for a short menu of things you can ask for directly.
The name stays yours. Click the name in the worker's header to rename it. Nothing the worker does ever changes its own name.
The words you'll see, in plain language
One sentence on what each thing is, and where to find it.
Approve and Reject
When its approval policy requires a check, the worker pauses and shows you the proposed action. Approve to allow that action or Reject to decline it. Either decision is applied once.
Where: A card in the chat, a message in Slack or Telegram, and a tap in the phone app.
Connections
Connect an account once and every worker can use it, with read and write access, until you turn it off for a worker in its settings (GitHub starts read-only; changing code still needs the repositories picked). A worker tells you what its job still needs and puts a Connect or Turn on button right in the conversation, so nothing has to be set up in a settings page first. With your Gmail or Outlook connected, ask any worker to go through your inbox and it offers to keep watching it: one tap starts watching, with every reply waiting for your OK.
Where: The card in the chat, Settings under Connections, and the account-wide Connections page.
Runs
Every piece of work is a run: what it did, step by step, how long it took, and what it cost.
Where: The Runs tab on the canvas beside the chat.
Receipts
Open a run and see where it got things: the pages it read, the company articles it cited and whether they were reviewed, what it sent to whom, and which engine did the thinking.
Where: Inside any run on the Runs tab.
Missions
Some jobs take longer than one sitting. A mission keeps going: it works, waits when it has to, and picks up where it left off, for hours or days, without you keeping the tab open.
Where: Type /plan and a goal in the composer, or the Missions tab.
Routines
A standing job with a name and a schedule. One worker can hold several: the 8am brief and the 6pm inbox sweep are two routines on one colleague, not two workers.
Where: The Routines tab, or just ask the worker to do something every day and it will offer to set one up.
Rooms
A group chat with several of your workers. Address one by name and only that one answers; leave the name off and the lead takes it. They hand work to each other in the open, the answer comes back to whoever asked, and a worker that is busy elsewhere is shown as busy and picks the message up when it is free. A worker can ask to open a room of its own, and you approve it like anything else.
Where: The Rooms band in the worker picker, and the Rooms window on the Theo Agent desk.
Apps
Ask a worker to show you something and it makes you a small live app: your inbox laid out by what happened to each email (what needs you, the drafts waiting — send, edit or discard them right there — what it replied and how, what it tidied, what it skipped), or a tracker over a list it keeps. Apps update on their own as it works. A worker reading your inbox keeps a log of what it did with each email (the subject, the sender, its one-line summary and the reply it sent, never the email itself), and you can switch that log off. An app's buttons only ask the worker, open a link or let you mark an item, so an app can never do more than the worker already may.
Where: The Live tab of the worker's computer, a card in the chat, and a window of its own on the Theo Agent desk.
Drafts waiting for your OK
Tell a worker that watches your inbox to let you approve replies first, and every reply it writes waits as a draft instead of going out. Send it as written, edit it first, or discard it, in any order and as many as are waiting, while the worker keeps working through the rest of your mail. A reply that needs something only you know is drafted with that part marked [[like this]], and Send stays off until you fill it in. You get one notification when drafts are waiting, a thread you discard stays with you until somebody writes again, and the worker can only make replies stricter on its own: letting them go out without you is your choice, in its settings.
Where: The Inbox app, on the worker's computer and in a window of its own on the Theo Agent desk. The notification opens it.
Tools it made
When a worker keeps doing the same kind of task, it can propose a tool for it: a name, what it is for and the steps it follows, built only from things it may already do. Nothing changes until you accept it, every step still asks when your Rules say so, and you can remove it any time.
Where: Settings, Advanced, under Tools it made.
Its desk
Every worker can open a computer of its own, with files, a terminal and a browser, and do the work there instead of describing it. Anything it leaves on its desk lands in your Drive.
Where: The Work tab on the canvas.
Watch and Take over
See the page the worker's browser is on, live. When it hits a login, a code or a payment step, it asks you to step in; you type into the real page and hand it back. It never sees or stores what you typed.
Where: The Browser tab under Work.
Work in a repository
Give it a code repository and it reads the code, makes the change, runs the tests and opens the pull request, with every command and edit on a timeline you can watch.
Where: The Work tab, and the Software Engineer starter.
Templates
Duplicate a worker you have tuned, or share its setup with your team or as a link. The goal, persona, tools and settings travel; accounts, passwords and memories never do.
Where: Settings, on the Identity tab.
Trust
Approve the same kind of action a few times in a row and the worker stops asking for that kind, and only that kind. One no resets the count. Deleting is never earned.
Where: Settings, at the foot of the Guardrails tab.
Evals
Before you change its engine, its instructions or its tools, replay its own past runs against the new setup and see what would change. Nothing is sent and nothing is done; it only thinks.
Where: The Evals tab, and Add to eval suite on any run.
Out of credits
Running out of credits is a pause, not a failure. The worker warns you once when a task looks bigger than your balance, lands what works first, and stops exactly where it is. A top-up resumes the same work.
Where: An amber card above the composer with Buy credits and Resume.
Brain
Which engine does the thinking. Let Theo choose per step, go economy, or pin one engine; pick how hard it thinks with Fast, Smart or Deep.
Where: The pill in the composer footer, or Settings on the Brain tab.
Knowledge and Memory
Your company wiki in front of it always or when it decides to look, and what it has learned about you: browse, edit, pin or forget any memory, and optionally let it learn how you write.
Where: Settings, on the Knowledge and Memory tabs.
The roster
The list of your workers shows what each one is doing right now, what it last said, and puts the ones that need you first.
Where: The worker picker and the Theo Agent list in the chat sidebar.
Why Theo Agent is in beta
Autonomous workers are a new kind of product: they plan and act on their own instead of waiting for your next prompt. We are testing Theo Agent with real users while we tune reliability, its channels and the safety rails around risky actions, so Theo Agent carries a Beta badge.
A worker can make mistakes. It can misread context, pick an imperfect step, or word a result differently than you would. Review anything important before you act on it, and keep approvals on for actions that reach the outside world. Every action is recorded on the Runs tab, and you can pause or delete a worker at any time.
How it stays safe
Actions come in two kinds: safe ones that stay inside your workspace (writing a note, building a deck, reading the web) and risky ones that reach the outside world, spend credits or delete things (sending an email or a text, posting in Slack, booking a meeting, pushing code). You decide when the worker must check with you:
Ask before risky actions
DefaultThe worker runs safe things on its own and pauses for your approval before anything that touches the outside world, spends credits or deletes something.
Ask before every action
Maximum oversight. The worker pauses before each step so you approve everything.
Run autonomously
No prompts. Best for trusted, low-stakes jobs and for anything that runs while you sleep, like a nightly engineer.
One approval can cover a whole plan. When a worker wants to take several risky steps in one go, it shows you the whole list with each step marked reversible or not, and one Approve runs them in order.
Your organisation can set rules over every worker. Teams admins can mark an action Always ask, so it pauses for a human even on a worker set to run autonomously, or Never allow, so it disappears from every org worker. Find it under Teams, Settings, General, Theo Agent policy.
Replies to people can wait for you. A worker that answers email, texts or Slack on your behalf can be set to send, to hold every reply as a draft for you, or to send only when it is sure. Settings, the On its own tab.
Every run is bounded. A step and time budget means a worker can never loop forever or run up a surprise bill, and a worker with nothing left to spend pauses rather than failing.
Reach it anywhere
Theo Chat
Every worker opens as a conversation inside Theo Chat, with a canvas beside it showing the work.
Telegram
Bind a worker to a chat and message it like a teammate; approve or reject with buttons; voice notes and photos understood.
Slack
Bind a worker to a channel; it answers in threads, reacts while it works, and posts Approve / Reject right there.
Its own email address and calendar
One click gives a worker an address, a calendar and contacts. People reply to it; it books meetings with real invitations.
Its own phone number
Claim a number and people can text the worker; it texts back from the same number and remembers each conversation.
Your own software
Every worker is reachable over the API and MCP, including an OpenAI-compatible endpoint.
The phone app
The roster, the transcript and Approve / Decline in one tap when a worker needs you.
Everything a worker says elsewhere shows up in its chat as a read-only thread card, so you never have to switch apps to know what it has been up to. Channels are set up in Settings, on the Channels and Connections tabs.
Channel details
Its own email address and calendar
Open the worker's Settings, find the Agent identity panel under Channels, and click Give it an address. That one step sets up its email address, its calendar and its contacts, with nothing to host or verify. The Name field is what people see on outgoing email and on invitations; you can also set a reply-to address, a default time zone and a signature.
When someone replies to a worker's email, the reply routes back to that worker and it answers from the same address; each correspondent is its own ongoing thread. Replying to the person who wrote is automatic; emailing anyone else follows the worker's approval posture. Files up to about 10 MB are attached directly; larger ones go as a secure download link.
The calendar is real: the worker checks its own free time before proposing slots, books meetings with actual attendees who get a normal invitation, reschedules, cancels and answers invitations sent to it. Booking counts as risky, so under the default posture it asks first; switch Let it change the calendar off and it can still answer “when am I free?” without being able to book.
Its own phone number (texts)
In Settings under Channels, open the Text messages panel, search for a number (by US area code if you like), and add it for a flat monthly fee; release it any time from the same panel and the charge stops. Anyone can text the worker and it texts back from the same number, remembering each person's conversation separately.
Texting a person back is automatic, even under the default posture; texting a different number or sending email still follows the worker's approval posture. STOP, HELP and START are handled for you with the standard carrier replies, and a number that opted out is never texted again. Undeliverable texts show up as a heads-up in your notifications.
Telegram
Bind a worker to a Telegram chat and it becomes a teammate you can message. Send it instructions, get its replies in the chat, and answer its approval prompts with Approve / Reject buttons. It shows a typing indicator while it works, and when a run cannot finish it says plainly what went wrong. Voice notes are transcribed, documents are read, and photos are understood along with your caption.
Without leaving Telegram: /agents lists your workers with a Bind button on each, /bind <name> and /unbind connect and disconnect one, /pause and /resume stop and restart it, /new starts a fresh session, and /status <name> shows what a worker is up to. The in-app chat and the Telegram chat are the same conversation.
Slack
Connect your Slack workspace once, then bind a worker to a channel from its Settings under Connections. Binding a public channel joins the app to it for you; a private channel needs an invite first. Anyone in the channel can @mention the worker; it reacts while it works, posts one status message it keeps updating, replies in the thread, and posts Approve / Reject buttons when it needs your sign-off. Once it has answered in a thread, people keep talking to it there without another mention.
A worker pointed at a channel listens to that channel; it never joins a conversation in a channel nobody bound it to. It reads files dropped into the thread, uploads real files back, attaches an Open button to anything it made, and knows whether it is talking to you, a teammate or a guest. Runs are billed to the account that connected the workspace. There are two Slack integrations on purpose: the worker's conversation channel, and Slack as a tool the worker can post through; the panel labels each so disconnecting one never breaks the other.
Several workers together
Attach helpers to a worker and it can hand a sub-task to one of them; the helper runs on its own budget and reports back, and its reply shows up in the same conversation with its name on it. A manager only delegates to workers you attached, never to a worker that would loop back to it, and never outside your organisation.
A room is the same idea in the open: you and several workers in one conversation, where you address a worker by name and workers can message each other, within limits you can see. Only the worker you named answers — leave the name off and the lead takes it — so several workers can work in one place without talking over each other. When one hands a part to another, it says so in a line you can read, and the answer comes back to the one that asked instead of being repeated at you. A worker that is busy somewhere else is shown as busy rather than interrupted, and it picks the message up on its next turn. A worker can also ask to open a room with teammates of its own; you approve that the same way you approve anything else, and the room appears with a door into it in the worker's own chat.
Want to know what a worker cost? Ask it: “what did this week cost?” It reports its own runs, outcomes and credits for the window you name.
Talk to a worker from your own software
Every worker is reachable over the OpenCharts API and MCP: send it a message on a thread, poll for the reply, the things it made or the approval it is waiting on, approve or reject over the API, and read its usage. There is an OpenAI-compatible endpoint too, so any OpenAI SDK can talk to a worker by setting the model to agent:<id>. Details live in the developer docs.
Related Articles
Workflow Builder
Persistent, multi-step Theo workflows that connect apps, run AI actions, dispatch sub-agents, and learn from feedback.
Theo Memory
Theo remembers your preferences, facts, and instructions across conversations for personalized AI.
Theo AI Overview
Overview of Theo AI — 11 chat modes, creation tools, generative UI, memory, and the sidebar mascot.
