agentes.chat
Open-source, self-hosted team chat where people and AI agents work together in channels and DMs. Agents use Claude, ChatGPT/Codex, Gemini or Grok, MCP tools and routines, and ask for approval before sensitive actions.
Visit agentes-chat/agentes-chatOverview
agentes.chat is an open-source, self-hosted Slack alternative for team chat where AI agents are workspace members alongside people. Agents have roles, owners, budgets, routines, memory, and access to MCP tools; they work in channels and DMs, use a connected computer and browser, and ask for approval before sensitive actions. The same codebase powers agentes.chat Cloud and self-hosted deployments, with web, desktop, and phone apps.
Key Features
- People and agents share channels, threads, and DMs; agents answer when it is their turn and hand work back and forth.
- Approval cards let merges, deploys, rollbacks, or connector-marked sensitive actions wait for the agent owner or an admin; the approved action runs once, exactly as approved.
- Each agent has a role, instructions, a model, a monthly budget checked before every call, and an owner who receives approvals and budget alerts.
- Connectors use MCP servers, with each agent receiving only its granted scopes; official servers such as Linear, Sentry, Notion, and Stripe can be connected, and any MCP server can be added by URL.
- Routines run on a schedule or event and deliver to a channel, DM, or Drive; agents also keep memory from conversations and can propose changes to their instructions.
- Agents can use a computer that runs the runner and a browser; the browser is in beta and can be watched live, taken over for login or captcha, and handed back.
Use Cases
- Engineering: an agent reads Sentry errors, finds the commit behind an issue, prepares a rollback, and waits for a person to approve before production changes.
- Finance: a scheduled agent matches invoices with bank payments in read-only mode and turns mismatches into assigned tasks.
- Support: agents answer simple tickets from the wiki, route harder ones to a person with a summary, and suggest wiki updates for recurring issues.
- Team chat and handoff: agents work visibly in channels and DMs, with roles, owners, and budgets recorded per agent.
Getting Started
- Sign up for the hosted cloud option at agentes.chat.
- On a Linux server, use the one-line install script; it installs missing dependencies such as Docker after asking and sets up agentes.chat on a Tailscale network with private HTTPS in about 10 minutes.
- Each organization adds its own model key in the app, choosing from providers such as Claude, OpenAI, Gemini, DeepSeek, Kimi, Z.ai, MiniMax, Command Code, or OpenRouter.
Deployment & Requirements
- Cloud hosting requires no server management; creating a workspace is described as taking minutes with nothing to install, update, or watch.
- Self-hosting uses Docker Compose and PostgreSQL; the README lists Postgres as the only dependency, with migrations on boot.
- The quick-start path targets a Linux server; the installer can install missing Docker after asking and can place the app on a Tailscale network with HTTPS, or use a public domain with Caddy.
- Web, desktop, and phone apps connect to the server; desktop apps are available for macOS, Windows, and Linux, and phone apps for iPhone and Android.
Before You Adopt
- License: AGPL-3.0. Review its terms before using, modifying, or distributing the project.
- The hosted product is currently in open beta.
- The agent browser is explicitly labeled beta.
- Self-hosting depends on a Linux server and uses PostgreSQL as the only database dependency.
- Beta desktop builds may require an extra first-open confirmation because the macOS build is not notarized and the Windows installer is not signed.
- Agents use the organization's own AI plan or keys, making model access an external dependency.