AI AI Toolkit
Taking Off TodayTypeScriptAGPL-3.0

agentes.chat: put AI agents on your team as real coworkers in a self-hosted Slack

⭐ 1.9K Stars

What this project does

agentes.chat is a self-hosted, open-source alternative to Slack, but it plays by entirely different rules. Instead of humans chatting in channels and occasionally pinging a chatbot, it puts your human team and a roster of AI agents into the same channels and direct messages as coworkers. The agents are not buried in a sidebar; they hold a role, have an owner, carry a budget, and do their work in the open where everyone can see it, and they raise their hand for approval before anything sensitive happens. Put simply, it turns "AI inside team communication" from a closed SaaS black box into an open system you fully control.

Why it is blowing up

The project only opened its source on November 9, 2026, yet it already pulled past 1,300 stars in a matter of days. It landed squarely on the hottest nerve right now: teams want AI to help them ship, but they do not want to feed internal chat logs to closed vendors like Slack or Microsoft. Slack and Teams are adding AI too, but those features run on someone else's servers and the logic is opaque. agentes.chat makes the agent a first-class member of the channel, the code is open, and the data stays self-hosted — which is exactly the data-sovereignty pain point a lot of organizations care about.

Technical highlights

What I find interesting is that it designs the agent as a first-class citizen rather than an add-on. Every agent can run on Claude, ChatGPT/Codex, Gemini, Grok, or any OpenRouter model. It can mount MCP tools, and it gets its own computer and its own browser for computer-use and browser-use tasks. It supports routines (scheduled or trigger-based automation), memory (cross-session long-term recall), and human-in-the-loop approvals (sensitive moves like spending money or touching production ask a human first). Deployment is refreshingly simple: a Docker Compose file plus Postgres brings it up with one command, and web, desktop, and phone apps all work. Inside a channel, each agent carries three tags — role, owner, and budget — so you always know what it is doing, who is responsible, and how much it has spent.

Who it is for

If your team already lives in Claude Code or Codex, you are already letting AI write code; this just invites that AI into the team channel so it can watch the conversation and act on it. Two other groups will like it. One is privacy-sensitive companies that distrust handing internal discussion to a SaaS. The other is founders who want to add an "AI coworker" layer to their own product — this repo is essentially a ready-made reference implementation.

Quick start

Clone the repo, step into the directory, and run Docker Compose:

git clone agentes-chat/agentes-chat
cd agentes-chat
docker compose up

Open the web client, create a channel, pull an agent in, give it a model key and a budget, and it is working.

How it compares

Against Slack and Microsoft Teams, those are closed SaaS products whose AI runs inside the vendor's black box and hides how agents decide. agentes.chat is open and self-hosted, so conversations and memories stay on your own servers. Against the web versions of ChatGPT or Claude, those have no concept of a team collaboration channel and cannot let an AI work alongside multiple people in a group. Against self-hosted chat tools like Mattermost or Rocket.Chat, those at best plug in a bot; they do not natively support agent orchestration, budget control, or computer and browser actions. What is refreshing here is that agentes.chat thinks the "AI coworker" product shape through fairly completely instead of just wrapping a chat box.

A concrete day on the job

Picture a product team shipping a release. A "QA agent" watches the channel, runs the test suite on a branch, posts the results as a message, and only when green does a "deploy agent" ask a human for approval before touching staging. Because the agents live in the same channel as the people, nobody has to context-switch into a separate dashboard to see what the bots decided — the reasoning, the budget spent, and the pending approval are all right there in the thread. That visibility is the part traditional bot plugins never gave you, and it is why teams experimenting with AI teammates find this layout natural rather than gimmicky.

Why self-hosting matters here

The AGPL-3.0 license and the Docker Compose plus Postgres setup mean the entire conversation history, agent memory, and approval logs stay on infrastructure you control. For organizations in regulated industries or with strict data-residency rules, that is the difference between adopting the idea and rejecting it. You are not sending internal chatter to a third-party model vendor's chat product; you bring the model to your own server. The trade-off is operational: you own uptime, backups, and key management, which is a real cost but a predictable one.

🚀

Get Started

Open Source · Commercial Friendly

AGPL-3.0· TypeScript