AI teammates that work with the permissions of whoever uses them

Kyron's team works on many fronts at once: schools, orders, paperwork, social media. They wanted AI teammates that could do real things in the company's systems, not just answer in a chat window.

With two conditions. Everyone should see and do only what they can already do today. And nothing should change without an explicit yes.

Where Coro comes from

We didn't invent Coro from scratch. It starts from an open-source project by Elie Steinbock, released under the Apache 2.0 licence and now called Rakazo.

That project gives every agent its own conversation, memory, scheduled routines and a computer to work on. A solid base. In August 2026 we started from there.

On top of it we wrote around 170 changes of our own: what it took to bring it into a real company.

What we added

• A signed Mac app that installs like any other app and updates itself.
• An iPhone app ready for each company's own server, with Google sign-in and switching between servers.
• Team agents: an admin turns an agent into a team agent, and every person gets their own copy.
• A software factory: one sentence starts agents that write the spec, build, review and merge the work. Or they stop and ask a person.
• Charts in the chat, on the web and on iPhone, even in the company's colours.
• Real-time voice calls: the voice answers straight away and hands real work over to the agent.
• Saved logins: the agent fills in a site's credentials without ever seeing the password.

Team agents: one copy per person

This is the piece that makes Coro fit for a team. A team agent isn't a shared account. Each person has their own copy, with its own conversation and memory.

The copy uses the same tools as the original, but acts under the identity of whoever is using it. If someone can't see a record in the back office, neither can their agent.

Every write goes through a confirmation card with just two buttons: confirm or cancel. There is deliberately no "always allow". And each person has a monthly spending cap, so nobody burns through everyone's budget by mistake.

Saved logins, invisible passwords

Sometimes an agent has to sign in to a website to do its job. When that happens, it asks for the credentials through a secure form inside the chat.

Coro stores them encrypted and types them into the browser on the agent's behalf. The agent only knows a login exists. Passwords and codes never pass through its conversation.

For Kyron: a server of its own, and Kyra

For Kyron we installed Coro on a dedicated server of its own. That's where Kyra, the team agent, works alongside a few other agents.

They are connected to Kyron's back office, and there confirmation is mandatory on every write: the card shows what is about to change before it happens. The team uses them from the browser and the apps.

The work ranges from questions about data to everyday tasks. One person on the team, for example, publishes Instagram posts by asking the agent: it prepares the image and the caption, then posts.

Why start from an open project

Rebuilding something that already exists, and works, would have been time badly spent. Starting from an open-source project let us focus on what matters to a company: permissions, confirmations, costs under control, apps on the team's phones.

And the code stays open. We keep our changes in files separate from the original wherever we can, so we can keep taking in improvements from the people who created it.

What changed

Kyron's team has AI teammates in the browser and in the apps. Each one works inside the permissions of the person using it, writes nothing without a yes, and has a spending limit.

It isn't one more chat to open. It's a different way of asking the company's systems for things.