Skip to Content

Self-hosted team chat to run your AI assistant

A messaging app you own, one account and one channel per purpose, alerts in the same place: your AI's control room.
November 15, 2026 by
Self-hosted team chat to run your AI assistant
Federico Solutions, Myrko Federico

Self-hosted team chat to run your AI assistant

Field: AI assistants, team messaging, self-hosting

For: SME owners who want to work with an AI assistant every day, and their IT person

Experience: several months running my AI assistant from a team messaging app hosted on my own server

An AI assistant installed on your server is useless if you can only talk to it from a terminal. It needs an interface the whole team knows how to use, on a computer as well as on a phone. The temptation is to plug it into the consumer messaging app everybody already has. I connected it to a self-hosted team chat that I own. Here is why, how to organise it, and the rules that stop it from becoming a back door.

Background

In my hands-on account of an AI agent in an SME, I described what my assistant did: keep my task list, look things up in my ERP, search the web, draft emails. That article was about what the assistant was allowed to do. This one is about where you talk to it from: the chat channel, which is both your control room and a way in to everything the assistant can touch.

The problem: running your AI from a tool you do not control

Plugging an assistant into a consumer or rented messaging app raises several difficulties:

  • Conversations go elsewhere: what you ask the assistant, and what it tells you about your customers or your figures, is stored by the messaging provider.
  • The channel becomes a remote control: anyone who can write to the assistant can give it orders, with all its rights.
  • The rules change without you: a rented messaging service can change its terms, its prices or its programming interface, and your assistant stops answering.
  • Everything gets mixed up: alerts, requests and answers land in the same thread, and the important information gets lost.

The solution: a chat you own

1. Host the messaging app yourself

I used Mattermost, an open source team messaging app that installs on your server, with a desktop app and a mobile app. There are other options, such as Matrix and its Element client. The essentials are the same: messages are stored in your database, on your server, and access goes through the company VPN.

2. Give the assistant its own account

The assistant connects with a dedicated account and its own access token, never with a person's account. That way you can clearly see what it writes, and you can cut its access by revoking that one token, without touching the team's accounts.

3. One channel per purpose

One channel for general requests, one for running the server, one for alerts. Each channel can be linked to a different agent, with its own rights: the one that manages the server does not answer in the channel open to the whole team. Threads keep each request and its answer together.

4. Choose who can talk to the assistant

The assistant only answers authorised people, and only in the intended channels. An explicit list is better than opening it to everyone: a message written by someone unexpected, or slipped into a shared document, must not become an order.

5. Bring alerts to the same place

The messaging app also receives notifications from your other tools through incoming webhooks: last night's backup result, a monitoring alert, an audit report. The owner sees in one place what the assistant says and what the machines say, without opening five consoles.

What I learned running it

  • The phone changes how it is used. Being able to ask about a project or my tasks from the mobile app, connected to the VPN, turned the assistant from a technical tool into a colleague you call on several times a day.
  • Keep alerts apart from conversations. Mixed together, automatic notifications bury the assistant's answers, and vice versa. A separate alerts channel that you scan at a glance solved the problem.
  • The channel does not replace the guardrails. Being able to write to the assistant from anywhere makes approval before any irreversible action all the more necessary: an order typed too quickly on a phone is still an order.
  • The tool can change, the principle stays. Today, for other uses, I talk to my work sessions through my AI provider's official apps, as I explain in the article on claudeVPS. The rule has not changed: an identified, reserved channel, and I know where the messages go.

Under the hood (for your IT person)

  • Deployment: Mattermost in a container with its PostgreSQL database, behind the reverse proxy, site URL reachable only through the VPN.
  • Assistant identity: a dedicated user account with a personal access token (or a bot account), rights limited to the channels it needs, token kept in the encrypted secrets file.
  • Routing: a channel → agent binding in the assistant's gateway, with a list of authorised users and a policy for direct messages.
  • Integrations: incoming webhooks for backups, monitoring and audits; slash commands if needed.
  • Operations: the messaging app is a service like any other: backed up, monitored, updated by hand.

What I take away

The chat you run your assistant from is not an interface detail: it is the way in to everything it can do. Hosted on your server, reserved for the right people and organised by purpose, it turns the assistant into a colleague you can reach anywhere without exporting your conversations. It comes into its own with an assistant that has a memory, which remembers from one request to the next, and it rests on a server closed to the internet and on well-kept secrets.

Going further

I design AI assistants and automations tailored to your team, with their channel, their rights and their guardrails. See custom applications and automations, or book a 20-minute discovery call.

The step-by-step guide

For your IT person, guide 4 in the Open Source Server Stack series, team messaging connected to the assistant (CHF 39), installs Mattermost and connects it to the assistant in 1 to 2 hours: deployment, dedicated account and token, channels per agent, list of authorised users, webhooks for alerts and commands. At the end, you type "summarise my tasks for today" on your phone, and the assistant answers.

The Guides Pack (CHF 149) brings together guides 0 to 8 of the series: a complete stack, documented and tested in production, from the secure server to the assistant's memory and monitoring. If you would rather have support, the assisted installation (CHF 597) installs the foundation with you in one or two video sessions, Complete Pack included, with 60 days of email support.

Sources

Tags: team chat, self-hosted messaging, Mattermost, AI assistant, AI agents, webhooks, digital sovereignty, SME

Managing server secrets without exposing them
Vaultwarden vault, sops and age encryption, in-memory decryption, rotation: no more plain-text passwords on disk.