AI NewsOpen sourceAnnouncement
Cumora is a team chat where AI agents join channels with humans
Cumora is a fifteen-day-old open source team chat where AI agents take part in the same DMs and group conversations as people, hold their own personas and memory, and can send and receive real email.
Image: GitHub
Why it mattersA shared roster removes the extra code teams keep writing around Slack bots, and shows what a workflow looks like when the agent is a permanent teammate rather than something called up once per thread.
Cumora is a new open source team chat where AI agents are members of the workspace with the same rights as any human account. It was created by GitHub user yetone on 17 August and now has 3,385 stars, 419 forks and about 190 commits on main. It is licensed MIT.
The README's own description reads: "Cross-platform team chat where AI agents are first-class participants alongside humans, same roster, same DMs, same group conversations, same Kanban board and calendar."
What that means in practice
Agents in Cumora hold a persona and a memory that persist across conversations. They can be added to a direct message or a group chat the same way a person is, claim work units on a shared Kanban board without two agents taking the same unit, and read and send real email through a Resend integration. Inbound email arrives through Cloudflare Workers. The chat, Kanban board and calendar are the shared space both people and agents work in.
The tool set the README lists for agents is small and concrete: bash, files, browser, email, and skills.
How it runs
Cumora is available on four platforms: an Electron desktop app, a progressive web app, an iOS build distributed through TestFlight, and an Android build users compile themselves. The Capacitor project is the base for the two mobile apps.
There are two ways to connect the agents to AI models. Cumora Cloud runs a managed per-agent Kubernetes pod on the project's own infrastructure, backed by the OpenAI Responses API. The bring-your-own-agent mode runs a daemon on the user's Mac or VPS and calls that user's own provider account, with Claude Code and Codex sandboxed by default. The README notes that the Cumora server never sees a provider key in that mode.
The reason so many teams write a Slack bot wrapper around Claude Code or Codex is that a chat platform assumes a person is on the other end of every account. Cumora reverses that: an agent is an ordinary member with a login, a mailbox and a share of the board. If it works, the wrapper code that most teams have written twice already is no longer needed, and the typical failures change from "the bot lost track of the conversation" to the ones humans have on a team, such as two agents picking up the same task.
That is also where the design is under-tested at fifteen days old. The atomic work claim, the persona memory and the email loop are the kind of basic features that only show their weaknesses on a real team over weeks. The interest is real, and the code is there to read. Try it on one channel with one agent before relying on it for a workflow.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
