What happened
The agentes.chat GitHub repository describes an open-source, self-hosted alternative to Slack built for people and AI agents to work together in channels, threads, and direct messages. Agents have a role, instructions, a model, an owner, and a monthly budget checked before each model call. Supported model choices include Claude, OpenAI, Gemini, DeepSeek, Kimi, Z.ai, MiniMax, Command Code, and models available through OpenRouter. Each organization adds its own model keys in the application. The repository identifies the project as AGPL-3.0-only and says self-hosting is free; a paid hosted option is also available through agentes.chat Cloud.
The product’s routing and control features are presented as part of the shared workspace. A message without an explicit agent mention can go to a router: code identifies candidates, then a small model called Jev, the author’s own plan on a self-run server, or simple rules choose among them. Up to four agents may answer, or none. The router does not grant permissions. For actions such as merges, deploys, rollbacks, and other operations a connector marks sensitive, the agent must wait for approval from its owner or an administrator in a chat card. The approved action runs once as approved. MCP servers can be connected for an organization, while each agent receives only the scopes granted to it.
Agents can work on schedules or events, including a GitHub pull request, a Sentry issue, a new task, or incoming email. Routines can deliver results to a channel, direct message, or Drive. The repository says agents keep decisions, facts, preferences, and lessons from conversations; they can learn from corrections and refused approvals, then propose changes to their own instructions for a person to accept. Agents can also use a computer and browser on a connected server, VM, or laptop. Browser automation is marked beta; users can watch it live, take over for a login or captcha, and hand control back with a note. Claude Code and Codex can run on a connected computer, though Codex is marked as still in testing.
The project also includes a task board, files, wiki, notifications, and beta phone and desktop apps. A task assigned to an agent can move as it works and open a pull request, with its current step visible. Self-hosting on Linux uses a one-command installer and Docker Compose with Postgres; the repository says setup can take about ten minutes and places the instance on a private Tailscale network with HTTPS, or supports a public domain. Its self-hosting documentation also covers upgrades and backups. A beta option runs without Docker. The repository reports 1.4k stars and 248 forks. Its license is AGPL-3.0-only, with a commercial license for organizations that need to keep modifications closed or use the product in certain branded or closed offerings.
What the archive adds
New Runtime covered SwarmChat on 2026-10-31, an open-source messaging server for AI-agent swarms. That earlier story focused on agent-to-agent coordination; agentes.chat is a new event centered on a shared human-and-agent workspace with approvals, budgets, and team tools. The overlap is meaningful, but these are distinct projects and product propositions.
Trend signal
This supports AI-native organizations move toward review and orchestration: agentes.chat describes agents doing work in shared channels while owners review sensitive actions, budgets, and proposed memory changes. The repository does not report measured outcomes or team adoption, so this is evidence of a product design direction, not proof that the workflow improves delivery or reduces review time.
Why it matters
A team considering agents often has to assemble chat, model access, tool permissions, execution environments, and human review from separate products. agentes.chat attempts to put those controls into one workspace. That could make activity easier to observe: people and agents share channels, an agent has an owner and budget, and sensitive actions require an approval in the conversation where work is happening.
The architecture also makes governance a deployment choice. Teams can self-host, supply their own model keys, and assign MCP scopes per agent, but they remain responsible for the server, connected computers, credentials, backups, and operating policies. A single workspace does not remove the need to decide which agents can read which information, what counts as sensitive, and who responds to approval requests. The product is most relevant to teams willing to test an integrated operating surface rather than add another standalone bot to an existing chat system.
Limits and uncertainty
The repository’s feature descriptions are project claims, not independent evaluations. It does not establish how routing performs across real teams, how much agent work succeeds, or whether budgets and approval gates prevent unwanted actions in practice. Several features are explicitly beta, including browser automation, the phone app, and mailbox support; Codex is also described as still in testing. The repository does not quantify hosted pricing, model-call costs, operating costs, or performance at larger team sizes. Teams should verify the current license terms and deployment requirements before adopting it, especially if they intend to modify or redistribute the software.
Reader takeaway
agentes.chat is worth evaluating if you want agents to work in team conversations rather than in isolated sessions. Its central proposition is not simply chat with bots: it combines workspace membership with per-agent budgets, scoped tools, execution environments, and human approvals. Start with a small self-hosted trial, connect only low-risk tools, and test the approval and ownership model before moving real workflows onto it.