Skip to content

Open-source Slack alternative puts AI agents in channels

agentes.chat is a self-hosted team chat where agents are workspace members with budgets, approvals and their own computers, and it gathered 1,358 stars in five days.

Draftship Editorial4 min read

A team chat where agents are members, not bots

agentes.chat is an open-source, self-hosted Slack alternative where AI agents join channels and DMs alongside people. The project's repository was created on 2026-11-09 and had 1,358 stars and 237 forks by 2026-11-14, five days later, according to the project's GitHub page. It is written mainly in TypeScript and released under AGPL-3.0.

The pitch is that agents are workspace members with a role, an owner and a budget rather than a bot bolted onto chat. They read the channel and answer when a message is theirs to handle, and work passes back and forth between people and agents.

How the agents actually work

The source says agents can think with Claude, ChatGPT and Codex, Gemini, Grok or any OpenRouter model. Each organization adds its own model key from Claude, OpenAI, Gemini, DeepSeek, Kimi, Z.ai, MiniMax or OpenRouter. Every model call is checked against the agent's monthly budget before it runs, and spend is recorded per agent.

When a message mentions nobody, it goes to a router. Code narrows the candidates, then a small model the source calls Jev picks among them, using the workspace's key, the author's own plan or simple rules. Up to four agents answer, or none, and the message states who chose.

Approvals are the safety layer. Merges, deploys, rollbacks and anything a connector marks sensitive wait for the agent's owner or an admin as a card in the chat, and the approved action runs once, exactly as approved.

Agents connect to MCP servers, with official ones such as Linear, Sentry, Notion and Stripe in one click or any server by URL, and each agent gets only the scopes granted to it. A Composio key extends that list to Google Drive, Calendar, Sheets and Docs, Slack, Microsoft Teams, HubSpot, Salesforce, Figma and others. Agents can also run commands on a connected computer, use a browser in beta, follow routines on a schedule or an event such as a GitHub pull request or a Sentry issue, and keep memory across conversations.

Self-hosting, cloud and licensing

Self-hosting is free under AGPL-3.0 with no licence key, and the source says that will not change. One command installs it on a Linux server, puts it on a Tailscale network with private HTTPS in about 10 minutes, and asks before installing Docker. Postgres with pgvector is the only dependency. There is also a beta without Docker, and web, desktop and phone apps connect to your own server.

agentes.chat Cloud is the paid hosted option and runs the same codebase. The source does not state Cloud pricing.

The licence terms matter if you plan to build on the code. AGPL-3.0-only applies, with a NOTICE that adds permission to distribute the phone and desktop apps through app stores and requires copies and modified versions to keep an attribution line in the About screen. Changing the code without publishing your changes, or reselling it under your own brand, needs a commercial licence from Codari Global, LLC. The setup CLI is MIT.

What it changes for app builders

Two patterns are worth copying even if you never run the product. First, per-agent budgets enforced before each model call, with spend recorded per agent, which turns runaway agent costs into a number you can cap. Second, approvals as chat cards for anything that touches production, so a deploy waits for a person and then runs once.

The beta Postgres connector applies the same idea to data access: agents query with SQL in a read-only transaction, one parsed statement, with time and row limits. If you are adding AI features to your own software, Draftship's guide to adding AI to your app covers the model side of that. Nothing in this announcement relates to Draftship, and Draftship is not affiliated with the project.

Limits and open questions

The source marks several parts as beta: the agent browser, the Postgres connector, agent mailboxes, the phone app, the desktop app, the no-Docker server, and imports from other tools. Codex is described as still in testing. The Mac desktop build is not notarized and the Windows installer is not signed, so the first open asks for an extra step.

The star and fork counts cover five days, which says something about attention and nothing about reliability. The source gives no benchmarks, no Cloud pricing, no security audit and no customer numbers.

What this means for building apps with AI

The design choices worth borrowing are unglamorous: a budget checked before every call, an approval that runs once, and scopes granted per agent rather than per workspace. Anyone shipping agent features will hit the same three questions. Who pays for the tokens, who approves the risky action, and what the agent is allowed to touch. agentes.chat answers them inside a chat product. If your app is the product, you answer them yourself.

Draftship is an AI app builder: describe an app in plain language and it writes the code, builds and hosts it, with a built-in database, user sign-in and hosting. Paid plans add private projects, a choice of AI model, source code access, custom domains and GitHub push. Nothing in the announcement says agentes.chat works with Draftship.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.