agentes.chat: a self-hosted Slack alternative where AI agents sit in your channels
agentes.chat (agentes-chat/agentes-chat) is an AGPL-3.0, self-hosted Slack alternative where people and AI agents share channels and DMs, with per-agent budgets and approval cards. One command installs it on a Linux server with Docker and Postgres.
agentes.chat (GitHub: agentes-chat/agentes-chat) is Slack-style team chat where AI agents are workspace members. Each agent has a role, instructions, a model, a monthly budget and a human owner, and it reads and posts in the same channels and DMs as people. Sensitive actions such as a deploy show up as approval cards in the chat. The code is AGPL-3.0 and self-hosting is free; agentes.chat Cloud is the paid hosted option, and Codari Global, LLC also sells a commercial licence. The GitHub repository was created on November 9, 2026 and had 1,491 stars and 266 forks on November 16.
Quick facts
- What it is: channels, threads, DMs, tasks, a wiki, a Drive and approvals, with agents as members. Web app, plus desktop (macOS, Windows, Linux) and phone (iPhone, Android) apps in beta.
- Server: one Bun process (web app, API, WebSockets, workers) plus Postgres 18 with pgvector, the only required dependency; Tailscale or Caddy handles HTTPS.
- Models: each organization brings its own key for Claude, OpenAI, Gemini, DeepSeek, Kimi, Z.ai, MiniMax, Command Code or OpenRouter, or any server speaking OpenAI Chat Completions or Anthropic Messages (Ollama, vLLM, LiteLLM).
- Licence and cost: AGPL-3.0-only with an attribution term; the setup CLI package is MIT. Self-hosting needs no licence key. Cloud is paid.
- Stack: Bun, Hono, Drizzle and PostgreSQL; React and Vite on the web; Expo for phones; Tauri for desktop.
- Versions (November 16): server v0.1.2 and desktop 0.8.10, both from November 15; 82 open issues, 117 open PRs.
How to self-host
The official route is one command on a 64-bit Linux server (amd64 or arm64); a Mac with Docker Desktop, OrbStack or Colima works for trying it:
curl -fsSL https://agentes.chat/install | shThe installer installs or starts Docker if needed (after asking), needs Docker Compose 2.23.1 or newer, fetches a pinned, checksum-verified Bun, then hands over to a setup wizard in /opt/agentes that writes .env with generated secrets and starts the containers. The first run downloads about 2 GB. To read it first, download install.sh and run it with --plan, which prints the installer's steps and runs nothing; the setup wizard has its own --dry-run.
By default the server joins your Tailscale network with a real certificate and no open port, so only your tailnet's devices reach it. First enable MagicDNS and HTTPS in the Tailscale console and generate a one-time auth key; later disable key expiry, or the server leaves the tailnet after 180 days. A public instance needs a domain, ports 80 and 443, and Caddy. Sizing has not been measured; the suggested start for a small team is 2 vCPUs, 4 GB RAM and 20 GB disk.
First setup: accounts, keys and agents
The first account to sign up becomes the owner and sign-up then closes; others join by email invitation or ask through a link for an admin to approve. On an internet-facing server, set OWNER_CLAIM_SHA256 so a stranger cannot take that first account. Without SMTP, emails go to the log and password reset is off.
The server needs no model key of its own. Onboarding asks the organization for one, stored encrypted and never sent to browsers; every model call checks the agent's budget first and records tokens and cost per agent. An agent can instead think with Claude Code or Codex on a connected computer, using a Claude or ChatGPT subscription signed in there (Codex is still in testing). The docs warn that answering others with a personal plan may break the provider's terms. An optional key for Jev, a fast model by TypeSafe, helps route messages nobody @mentioned. Agents that run commands need a computer running the runner: the server itself (the setup offers this) or a VM or laptop you connect.
What stands out
- A router, not a free-for-all. A message with no @mention goes to the router: code picks candidates, then Jev, the author's own plan or simple rules choose up to four agents, or none, and the message shows who picked. The router never grants permissions.
- Approvals in the conversation. Merge, deploy, rollback and anything a connector marks sensitive wait for the agent's owner or an admin; the approved action runs once, as approved. An agent's approvals can be loosened, down to a Never ask setting.
- Scoped tools. An organization connects an MCP server once, and each agent gets only the scopes granted to it. A beta Postgres connector runs read-only SQL with time and row limits.
- A computer and browser per agent. Agents run code on a connected machine and drive a browser (beta) you can watch live and take over for a login.
- Written-down decisions. docs/adr holds over 180 architecture decision records.
How it was built with Claude
The first commit on main is dated October 14, 2026, nearly four weeks before the GitHub repository was created. By November 15, main had 9,314 commits (2,471 of them merges); GitHub's contributors list showed 27 accounts, almost all work by Hélio Costa (6,639 commits) and wesley (2,527 under two spellings). Of those commits, 7,725 carry a Co-Authored-By trailer naming Claude, about 6,800 naming Claude Opus 5.5 and about 890 Claude Fable 5.2. Another 38 were authored by a GitHub account named claude.
The repo's CLAUDE.md lists non-negotiables: Postgres as the only required dependency, an org_id on every tenant table, keys never reaching a client, immutable migrations and an ADR for every architecture decision. Lint, typecheck, tests and a boundaries check must pass before work counts as done.
Try it with your AI. Give this to your own AI assistant for a cautious first install that stops before anything changes the machine.
Help me self-host agentes.chat (github.com/agentes-chat/agentes-chat) on my server.
Follow its SELF_HOSTING.md only.
Server: [Linux distro and version, amd64 or arm64]
Access: [private on Tailscale / public domain: your-domain]
First check, and report before doing anything else:
- uname -m and the OS; stop if it is not 64-bit Linux or macOS.
- docker compose version; it must be 2.23.1 or newer.
- Free RAM and disk against the project's starting point of 2 vCPUs, 4 GB RAM, 20 GB
disk.
Then:
1. Download the installer without running it: curl -fsSLo install.sh
https://agentes.chat/install
2. Summarise what install.sh does, then run sh install.sh --plan and show me the plan
(it covers the installer only; the setup's own plan is its --dry-run). Wait for my
yes before the real install.
3. If I choose Tailscale, list the console steps (MagicDNS, HTTPS, a one-time auth key)
and let me paste the key into the setup myself. Never ask me to show it to you.
4. Before the setup makes this server the agents' computer, explain what that lets
agents run here and ask me.
5. If the server is reachable from the internet, tell me how to keep the first account
for myself (OWNER_CLAIM_SHA256) before the address goes live; I generate and keep the
owner token myself. Confirm ALLOW_SIGNUP stays false.
6. Explain TELEMETRY=off and FEEDBACK_RELAY=false in .env, and ask whether I want them.
Rules: show me every command that installs software, edits .env or restarts containers,
and wait for my yes. Do not read or print .env secrets or model keys. Do not delete
files, volumes or containers.Known limits
- Young and moving fast. The GitHub repository is a week old; the desktop, phone, agent browser, mailboxes, Postgres connector and imports are beta, and open issue #564 reports the latest Docker image lagging a merged fix.
- AGPL terms. Per the project's COMMERCIAL.md, you can run it for your company and charge clients for installing it without asking. Offering a changed version to others requires publishing your changes under the AGPL and showing users where to get the source; NOTICE requires keeping "Based on agentes.chat by Codari Global, LLC" in the About screen. The name and logo are trademarks, so a fork needs its own. Contributions require signing a CLA.
- Agents act with real access. SECURITY.md calls a prompt that makes an agent misuse its access a known limit of agents. ADR 0117 tells agents to treat web pages, files and command output as data, and masks credentials in runs that read them. In the default install, agents share one OS user on their computer, so the OS does not keep them out of each other's sign-in folders; the --system install gives each agent its own user.
- What leaves your server. A daily anonymous usage report (counts, never messages, files or names) goes out through Cloudflare unless TELEMETRY=off. The feedback form sends the writer's message, name and email to the agentes.chat team unless FEEDBACK_RELAY=false.
- Unsigned desktop builds. In beta, the Mac app is not notarized and the Windows installer is not signed, so the first open needs a manual override.
Related on claudefox
- PhoneBox turns an old Android phone into a desk panel and control deck for Windows, Linux and macOS
- Pinky lets AI agents use your iPhone or Android phone, with approvals and blocked apps
- Agent Autoloop: run Claude Code and Codex in the background and get a card when they need you
- BAMA: Claude Code and Codex skills that plan, design and build native macOS apps
Problems and feedback
Questions go to GitHub Discussions, the project's support channel; confirmed bugs and feature requests go in issues. Report security problems privately through GitHub's vulnerability reporting.
FAQ
Is agentes.chat free to self-host? Yes. The code is AGPL-3.0 and self-hosting needs no licence key; agentes.chat Cloud is the paid hosted version of the same codebase.
Do I need my own AI API key? The server needs none. Each organization adds its own key in Settings, or an agent can think with Claude Code or Codex on a connected computer using the plan signed in there.
Can I use local models such as Ollama? Yes. The Other provider option takes any server that speaks OpenAI's Chat Completions or Anthropic's Messages, such as Ollama, vLLM, LM Studio or LiteLLM.
Can I run it without Docker? Yes, in beta: agentes-server is one program with its database built in, and the desktop app can run agentes.chat on a Mac or Linux computer.
Source: GitHub: agentes-chat/agentes-chat, agentes.chat README, agentes.chat SELF_HOSTING.md, agentes.chat CLAUDE.md, agentes.chat SECURITY.md, agentes.chat COMMERCIAL.md (licensing), agentes.chat NOTICE, agentes.chat CONTRIBUTING.md, agentes.chat SUPPORT.md, agentes.chat docker-compose.yml, ADR 0060: one-command install, ADR 0117: injection hardening, ADR 0180: dual licensing and CLA, ADR index, agentes.chat releases, agentes.chat commit history, agentes.chat open issues
This story first appeared in the claudefox daily briefing for November 16, 2026.