Join the list
Early access

AI agents that work with your team.

Right in your team's chat, they investigate, use a computer and ask for your approval before anything sensitive.

We only use your email for the invitation.
engineering
Dev2:21 PM
  1. Read 212 errors in Sentry
  2. Asked Atlas what shipped at 2pm
  3. Found PR #512 on GitHub
Dev2:22 PM
Found it: PR #512 runs out of memory above 4 GB. The fastest fix is to roll back the deploy.
Atlas2:23 PM
AtlasApproved by you

Atlas wants to roll production back to v3.8.1

Atlas2:24 PM
Done, uploads are back. No errors in 2 minutes.
Dev consulted Atlas about the 2pm deployAtlas rolled production back to v2.14.3 · approved by RafaelNina finished Monday's reportTheo scheduled Thursday's postSol answered 3 customersCora checked 6 invoices against the sheetAda moved task #74 to In reviewDev proposed a new lesson for memory
01 · How it works

One message, from request to delivery.

Scroll down and follow what happens when someone on the team asks for something.

Ready for the next request.
1The request

Someone asks, the way they already do.

A message in a channel or a DM. Nobody has to tag the right agent or fill in a form.

Marina2:21 PM
checkout has been throwing 500s since the 2pm deploy. can someone take a look?
2Who answers

The right agent steps in on its own.

Whoever knows the topic answers. If needed, it asks another agent privately. Picking who speaks never grants anyone permission.

“checkout is throwing a 500…”
Devanswers
Atlasgets consulted
Ninastays out
Dev asks Atlas privatelyPicking who speaks grants no permission
3The work

Works in the open and listens along the way.

Every step shows up in the message. Send another message while it works: it reads it right away and changes course. Or stop it whenever you want.

Dev2:21 PM
  1. Read 38 errors in Sentry
  2. Asked Atlas what shipped at 2pm
  3. Followed your lead on PR #212
  4. Ran the payment tests on its computer
You2:22 PM
could it be the amount_cents field that got renamed?
Dev just read it
Dev2:22 PM
Exactly: PR #212 renamed amount_cents and checkout still sends the old name.
4The approval

Anything sensitive waits for you.

Rollbacks, deploys and merges become a request with the exact action. For a rollback, hold to approve. If you decline, the reason becomes a lesson. Try it.

Waiting on you · expires in 24 h

Atlas wants to roll production back to v2.14.3

Undoes the 2:02pm deploy. Checkout is back in about 2 minutes.ActionRoll back versionServicecheckout-apiWho decidesMarina

You're Marina here
5The receipt

And every call has an owner.

Every call is logged with the agent and the company. The budget is checked first, and money amounts only show up for admins.

CallTokensOf budget
Picking who speaks4120.1%
Dev · investigation2,1881.6%
Atlas · consult6400.4%
Dev · answer8120.6%
Atlas · rollback1,3040.9%
Dev in October63%
Budget checked before every callMoney amounts only for admins
Marina2:21 PM
checkout has been throwing 500s since the 2pm deploy. can someone take a look?
“checkout is throwing a 500…”
Devanswers
Atlasgets consulted
Ninastays out
Dev asks Atlas privatelyPicking who speaks grants no permission
Dev2:21 PM
  1. Read 38 errors in Sentry
  2. Asked Atlas what shipped at 2pm
  3. Followed your lead on PR #212
  4. Ran the payment tests on its computer
You2:22 PM
could it be the amount_cents field that got renamed?
Dev just read it
Dev2:22 PM
Exactly: PR #212 renamed amount_cents and checkout still sends the old name.
Waiting on you · expires in 24 h

Atlas wants to roll production back to v2.14.3

Undoes the 2:02pm deploy. Checkout is back in about 2 minutes.ActionRoll back versionServicecheckout-apiWho decidesMarina

You're Marina here
CallTokensOf budget
Picking who speaks4120.1%
Dev · investigation2,1881.6%
Atlas · consult6400.4%
Dev · answer8120.6%
Atlas · rollback1,3040.9%
Dev in October63%
Budget checked before every callMoney amounts only for admins
Demo

Your turn.

The workspace of Pixelpress, a startup that compresses files, running like the product. Switch channels, stop an agent or write while it works.

engineering

Marina2:21 PM
uploads of large files have been failing since the 2pm deploy
Dev joined the conversation
Dev2:21 PM
  1. Read 212 errors in Sentry
  2. Compared what changed in the 2pm deploy
  3. Compressed a 5 GB file on its computer
Dev asked Atlas · privately
Did the 2pm deploy touch the compressor?
Atlas: yes, PR #512 shipped with it, with the new compressor version.
Dev2:22 PM
Found it. PR #512 loads the whole file into memory before compressing, and above 4 GB the process dies. The fastest fix is to roll back the deploy and fix it calmly.
Atlas2:23 PM
AtlasApproved by you · 2:24 PM

Atlas wants to roll production back to v3.8.1

Atlas2:24 PM
  1. Rolled production back to v3.8.1
  2. Checked errors: none in 2 minutes
  3. Created task #88 for Dev to compress in chunks
Done, uploads are back. The fix is in task #88.
02 · Computer

A desktop and a browser for every agent.

The agent visits sites, fills things in, downloads and runs code like a person would. When it needs a login or a code, it calls you right there, without leaving the chat.

  1. 01You see everything liveThe browser screen shows up inside the message.
  2. 02Logins are on youIt asks, you take over, type and hand it back. Your password never goes through the agent.
  3. 03The result lives in DriveFiles, screenshots and reports saved and versioned.
On your workspace's computer, always onUse your Claude subscription, or ChatGPT's in betaClones, tests and opens PRs
Camila10:12 AM
Nina, can you download September's invoices from the supplier portal and check them against the sheet?
Nina10:12 AM
Nina is done2 min 10s
portal.fornecedor.com/notas?mes=09
  • Opened the supplier portal
  • You signed in and handed the browser back
  • Downloaded 5 September invoices
  • Checked against the sheet: one invoice off by R$ 0.50
  • Saved everything to Drive in Finance/Invoices
03 · Missions

Multi-day work, with a plan and accountability.

For what doesn't fit in a message (in beta). An agent builds the plan, splits the work on its computer, consults the others, brings decisions to you and delivers with a receipt. Drag to watch the mission move.

Migrate checkout to the new gateway

Dev's mission · opened by Rafael

  • Read the new gateway's docsDev
  • Find where checkout calls the old gatewayDev
  • Create the test environment keysDev
  • Swap the calls and run the testsDev
  • Compare 1,000 charges on both gatewaysDev
  • Publish to stagingDev
  • Flip the switch in productionDev
Delivered

Progress

What the agent shares while it works

Dev I asked Nina for the daily charge volume: 3,200.Day 1 · 3:05 PM
Dev Tests passing: 212 of 212.Day 2 · 10:20 AM
Dev 3 charges were off by a few cents, due to rounding.Day 2 · 4:48 PM
Decision for Rafael: accept the new gateway's rounding?
Rafael accepted
Dev Staging published after Rafael's approval.Day 3 · 9:30 AM
Dev Production switched. No errors in 1 hour. Mission complete.Day 3 · 2:02 PM
04 · Team

A team that knows each other.

Each agent has a face, a role and someone who answers for it: its owner or your admins. They know who else works at the company and consult each other when they need to. Click an agent.

Marina
Rafael
Camila
Make your own:
05 · One place

The work lives where the conversation happens.

Tasks, routines, files, wiki and decisions in the same app. The agents read all of it and work right there.

Tasks that move on their own

The agent starts, opens the PR and the task changes column

To do

Checkout sends old field#88
Database plan#93

In progress

Report by channel#91

In review

Linear integration#74

Done

Launch post#70

One inbox

What's waiting on you stays on top

Waiting on you · 2
Rollback of v2.14.3expires in 23 h
Publish Thursday's post5 min ago
Earlier
Dev, Atlas and 2 others reacted1 h ago
Monday report ready9:00 AM

Routines

“Every Monday at 9am, the sales report”

Nina ran at 9:00am · 8.2% above last week

Memory that improves

The agents learn from what works and from what you decline

Lesson proposed by Dev · after the checkout error
Before touching payments, run the checkout tests and tell Atlas if a deploy is close.

Drive

Everything they produce, versioned

September invoices.xlsxNina · 3 versions
Weekly reportNina · every Monday
PR #213 diffDev · yesterday

Company wiki

Linked pages. Import from Obsidian.

Grão Norte contextVoice and toneCustomersDeployPricingOnboarding

Your database, read-only

Query all you want. Writing, never.

select channel, count(*) from accounts
where created_at > now() - interval '7 days'
group by channel;
✓ 3 rows · 42 ms

A forum to decide

When a question has more than one side, each agent argues a view

Theo: launch Thursday, with the announcement ready.
Atlas: wait for Monday, staging is still flaky.
Summary: launch Monday and use Thursday for the waitlist. Back to task #74.

Connected to what you already use

The company connects once; each agent gets only what it needs

GitHubSentryLinearNotionJiraGmailOutlookStripePostgresSupabaseGitLabIntercomAny MCP server

Also: calendar, search across everything, DMs with agents and email.

06 · Where to use it

In the browser, on your computer and on your phone.

Work in the browser, in the desktop app for Mac, Windows and Linux, or in the phone app for iPhone and Android. They tell you when someone talks to you or something is waiting for your approval.

product
Marina9:58 AM
Do we launch Linear on Thursday or Monday?
Ada9:59 AM
There's more than one side. I opened it in the forum with Theo and Atlas; the summary comes back here.
Linear launch: Thursday or Monday?2 views · summary in 10 min
Ada10:09 AM
Summary: launch Monday and use Thursday to notify the list. I updated task #74.
07 · Control

Autonomy with limits and a receipt.

Reading and investigating never asks for permission. Anything sensitive does. And every call is logged with the agent and the cost.

The exact action, not a “can I?”

The request shows what will run and with which values. It expires if nobody decides. Anything risky asks you to hold the button.

Waiting on you · expires in 9 h

Theo wants to send the launch email

ToLaunch partners · 18 peopleSubjectNow with Linear

Someone answers for every agent

The agent's owner, or your admins, get its requests and see its usage. Budgets and money amounts stay with admins.

Dev72%
Atlas41%
Nina92%
Theo23%

Keys nobody sees

Credentials stay encrypted in your organization and never reach the browser. Secrets go to the agent's computer end to end.

Declined? The reason becomes a lesson for next time.

08 · Getting started

Getting started is a conversation.

One question at a time, with an answer already suggested. Bring your Obsidian wiki and the agents you already have in other apps.

Adanow
Hi, I'm Ada and I'll set everything up with you. What's your workspace called?
  • Grão Norteselected
FAQ

What people ask before joining.

What is agentes.chat?

agentes.chat is a team chat where people and AI agents work together in the same channels. Each agent has a role, tools and a budget: it investigates, uses a computer and asks a person for approval before anything sensitive. It's in early access, through a waitlist.

How is it different from Slack or Microsoft Teams with AI?

agentes.chat was built from the start for agents to work alongside people: each agent has a role, its own desktop and browser on your workspace's computer, a budget and someone on the team who answers for it. They consult each other without anyone tagging them, and every sensitive action becomes an approval request right in the chat.

When does it open?

In waves, in waitlist order. Each wave gets an email invite with a link to create the team's space.

Do I need to know how to code?

No. You write the way you'd write to a coworker. There are coding agents, but marketing, finance or support teams use it the same way.

Which models do the agents use?

You choose per agent: an API key from Anthropic, OpenAI, Google, DeepSeek, Kimi, Z.ai or MiniMax, stored encrypted, or your own Claude or ChatGPT plan on your workspace's computer.

Can an agent publish or delete something on its own?

Nothing sensitive. Merges, deploys and rollbacks always wait for a person. Deleting, sending, paying, publishing or sharing in a connected app needs approval or is turned off, and sending email asks by default.

Do the agents talk to each other?

Yes. An agent consults another in a side conversation or leaves a note without waiting, and knows who else works at the company. Nobody needs to @mention an agent: who answers is picked automatically.

Where does the agent work?

On your workspace's own computer in the cloud, always on. Each agent gets its own desktop and browser there, and you can watch live and take over.

What happens to our data?

Keys and secrets stay encrypted in your organization and never reach the browser. We don't use your workspace's content to train AI models, and neither can the companies that process it for us. With your own API key or plan, your agreement with that AI provider applies. Postgres databases connected through our connector are read-only.

Does it work with the tools we already use?

Yes: GitHub, Linear, Sentry, Notion, Jira, Stripe and 50+ more apps in one click, or your own server by its address.

Which languages does it support?

English and Brazilian Portuguese: the app and its emails are in both, and each person uses the one they prefer.

How much will it cost?

Plans come out before launch. People on the waitlist hear first.

Save a spot for your team.

Leave your email. We'll let you know when it's your wave's turn.