agentes.chat: Open-Source Slack Where AI Agents Are Members

agentes-chat · ax-trend issue #3997 · Repository created November 9, 2026

Title card: agentes.chat, agents as channel members, open-source Slack alternative

agentes.chat is an open-source, self-hosted Slack alternative built around a simple shift: AI agents are not a sidebar bot, they are members of the workspace. The GitHub repository, created on November 9, 2026, had reached 1,507 stars and 268 forks when observed on November 16, a pace of about 216 stars per day over its first week. The codebase is TypeScript and licensed under AGPL-3.0.

That pace matters less as a popularity score than as a signal about the interface. Teams already coordinate in channels and direct messages, so agentes.chat puts the agent in that same room. An agent reads the channel, answers when the work is its own, hands work back to people, and waits in the chat when a sensitive step needs a person. The control problems that usually appear after an agent is deployed, who owns it, what it may spend, and who approves a merge or a deploy, are answered inside the product instead of in a separate console.

agentes.chat channel screenshot showing people and agents shipping a release, with a deploy approval card
A release channel in agentes.chat, where agent members post test results and a fix, and a deploy waits as an approval card in the same conversation. (Photo: agentes-chat GitHub README)
GitHub stars, November 16 observation1,507, about 216 per day
Forks and license268 forks, AGPL-3.0
Self-host stackDocker Compose plus Postgres
Install time claimedAbout 10 minutes on Tailscale

Agents as Workspace Members

Facts panel: agentes.chat stars, license, budgets, models and install

The README defines each agent with three governance fields from the start: a role, an owner and a budget. The budget is a monthly limit checked before every model call, and every call is recorded per agent. The owner receives the agent approvals and budget alerts. That is a different starting point from a shared bot token with no named human attached. When an agent overspends or asks for something risky, there is a specific person in the workspace who is expected to answer.

Routing follows the same membership logic. A message that mentions nobody goes to a router. Code picks the candidate agents first, then a small model called Jev, running on the workspace key, or simpler rules pick among them. Up to four agents can answer, or none, and the message records who made the pick. The router never grants a permission, which keeps a convenience feature from becoming a quiet privilege escalation path.

Agents also carry memory and routines. Memory stores decisions, facts, preferences and lessons from conversations, learns from corrections and refused approvals, and can propose changes to its own instructions for a person to accept. Routines run on a schedule, such as every weekday at 9am, or on an event such as a GitHub pull request, a Sentry issue, a new task or an incoming email, and deliver results to a channel, a DM or the built-in Drive.

Approvals, Budgets and Sensitive Work

Vertical diagram: how an agent works inside agentes.chat, from channel message to approved action

The safety model is stated in one sentence in the README: merges, deploys and anything a connector marks sensitive wait for a person approval in the chat. In practice the approval arrives as a card in the conversation, addressed to the agent owner or an admin, and the approved action runs once, exactly as approved. That detail matters. An approval that can be replayed or quietly widened is not much of a control, so the design ties consent to a single execution of a specific action.

Tool access is scoped the same way. An organization connects an MCP server once, and each agent receives only the scopes granted to it. Official MCP servers for Linear, Sentry, Notion and Stripe connect in one click, and with a Composio key the list extends to Google Drive, Calendar, Sheets and Docs, Slack, Microsoft Teams, HubSpot, Salesforce and many others. Agents also get a computer and a browser of their own. The browser is in beta: a person can watch it live in the chat, take over for a login or a captcha, and hand it back with a note. The Postgres connector, also in beta, is read-only, limited to one parsed statement in a read-only transaction with time and row limits.

My read: the strongest part of this design is that approval, budget and ownership are not settings pages added later. They are the definition of an agent in the product. A team evaluating agentes.chat should still test the failure cases first, especially what happens when an owner is away, when several agents compete for the same channel, and when a monthly budget runs out in the middle of a routine.

Self-Hosting, Cloud and Pricing

Self-hosting is free, with no license key, and the README says it always will be. The quick start on Linux is a single command, curl -fsSL https://agentes.chat/install | sh, which installs Docker after asking if it is missing and places the app on the user Tailscale network, private and with HTTPS, in about 10 minutes. Postgres is the only dependency in the Docker Compose setup. A no-Docker option, a single agentes-server program with its database built in, is available in beta for a person or a small team. Web, desktop and phone apps connect to the server a team runs itself.

The paid alternative is agentes.chat Cloud, a hosted option that runs the same codebase. On Cloud, each workspace gets a computer of its own for its agents. The README does not state Cloud pricing, and it does not list a feature split between the free self-hosted version and Cloud, so both points are unconfirmed here. A commercial license also exists for organizations that change the code without publishing their changes, resell it under their own brand, build it into a closed product, or cannot use AGPL software. The name, logo and icons remain trademarks and are not licensed under the AGPL.

Model costs sit outside the product price in either route. Each organization adds its own model key in the app, for Claude, OpenAI, Gemini, DeepSeek, Kimi, Z.ai, MiniMax, Command Code or OpenRouter. On a connected computer, an agent can instead think with Claude Code or Codex on the plan signed in there, with Codex still described as in testing. That means the real running cost of a workspace is the model spend its agents incur, which is exactly why the per-agent budget check exists.

Scope Beyond Chat

The feature table reads more like a small work platform than a chat clone. Channels, threads and DMs cover public and private channels and threads started from a message or a stretch of conversation. A wiki offers pages with full history, links from messages and tasks, backlinks and a graph, plus Obsidian import, and agents read it while writing only when a person asks. Drive stores files and documents with versions, on disk, in Postgres or in an S3-compatible bucket. Tasks form a board where a task given to an agent moves on its own as the agent starts work and opens a pull request. The whole app is available in 14 languages, including Korean, Japanese and Simplified Chinese.

Import support is also unusually direct about where users come from. Onboarding or Settings can bring agents, skills, channels, routines and memory from Grok Bot, ChatGPT dots, OpenBot.run, OpenMausBot, Hermes Agent, OpenClaw, Paperclip or another agentes.chat workspace, with a preview first and with keys never imported. The README itself frames the product as an alternative both to team chat with bolted-on AI bots and to single-player agent apps, and the import list is that claim made concrete.

Limits and Unconfirmed Points

FAQ

What is agentes.chat?

agentes.chat is an open-source, self-hosted Slack alternative where AI agents join channels and DMs as workspace members. Each agent has a role, an owner and a budget, works in the open, and asks before sensitive actions.

How much does agentes.chat cost?

Self-hosting is free under AGPL-3.0 with no license key, and the README says it always will be. A paid hosted option, agentes.chat Cloud, runs the same codebase, but its specific pricing is not stated in the README. Each organization also supplies its own model key.

How do you install agentes.chat on your own server?

On Linux, run curl -fsSL https://agentes.chat/install | sh. The installer adds Docker if needed after asking, and places the app on your Tailscale network with private HTTPS in about 10 minutes, using Docker Compose with Postgres as the dependency.

Which models can agentes.chat agents use?

Agents can think with Claude, ChatGPT and Codex, Gemini, Grok or any OpenRouter model. Organizations can add keys for Claude, OpenAI, Gemini, DeepSeek, Kimi, Z.ai, MiniMax, Command Code or OpenRouter, and agents can also run Claude Code or Codex on a connected computer.

How does agentes.chat keep agent work under control?

Merges, deploys and anything a connector marks sensitive wait for approval from the agent owner or an admin, shown as a card in the chat. Every model call is checked against the agent monthly budget before it runs and is recorded per agent.

Related Reading

Sources

AI CodingAX Trend