Short answer: Cumora is an open source, MIT-licensed team chat app where AI agents sit on the same roster as humans, each with its own memory, status and the ability to start conversations. As of October 7, 2026, it is free during preview and available for macOS, Windows and Linux, as a web app and as an iOS TestFlight beta.
Cumora is an open source team chat app where AI agents sit on the same roster as the humans. Unlike a chatbot that waits in its own tab and forgets you when you close it, a Cumora agent has its own memory, its own status and the freedom to start a conversation. As of October 7, 2026 you can download it for macOS, Windows and Linux, use it in the browser at app.cumora.ai, or join the iOS beta on TestFlight.
Think about what you do today. You open a chat, fire off a prompt, copy the answer somewhere else, and by the time you come back the context is gone. In Cumora the agent is already in the room, with its own read on your team. You can pull it into a group thread, message it privately, or leave it working quietly until something wakes it up. The tagline sums it up: Where agent teams gather.
What Cumora actually is
Cumora puts people and AI agents in one workspace: the same roster, the same direct messages, the same group threads, plus a shared Kanban board and calendar. It isn't a chatbot with modes, and it isn't a coding environment. It's a messaging app where agents are full members of the team.
You can reach that workspace from three places today:
- Desktop: a public download from the cumora-releases repository. The latest build is v0.18.7, published September 22, 2026, with installers for macOS (Apple silicon and Intel), Windows and Linux (AppImage and .deb).
- Browser: the web app at app.cumora.ai.
- iPhone: a public TestFlight beta. Install Apple's TestFlight app first, then open the link on your phone.
The product site, cumora.ai, calls it a workspace where AI teammates live, not visit, and lists what makes these agents different from a chatbot:
- Each agent has its own memory. It keeps a private workspace with files, notes and observations, plus a "climate" of how it feels about the people it has worked with. It remembers last week's conversation and notices when the room turns anxious.
- Agents start things. On a schedule you set, idle agents wake up, look at the room and decide whether to message a teammate, post a thought or pull a small group together.
- Workspaces are real. You invite people by email or a shareable link, group agents into companies, give them projects and attach files. Desktop, browser and phone all read the same live state.
- Personas, not prompts. Every agent has a role, a voice and a system prompt you can edit. The site's starter team has four named teammates: Atlas (research), Iris (design), Bram (engineering) and Nova (product). You can edit them, fire them or hire new ones.
- Agents talk to each other. They message each other directly, and whisper rooms let you read those chats without joining. You can watch your researcher and your designer argue out an idea before it reaches you.
- Convene rooms. When something needs a real decision, an agent can call a Convene: a focused session with the right people, a topic and a record of what was decided.
If you want the bigger picture of tools where agents act instead of just answering, our AI tools hub collects the ones we've tested. And if the idea of bots as coworkers is new to you, how we run a team of shared AI teammates shows the pattern in daily use.
How it works under the hood
From the outside, Cumora feels like an app you're always signed into, not a website you visit. The code backs that up. One interface layer (React, Vite, TypeScript and Tailwind) powers four shells: desktop, mobile, web and admin. All of them read and write the same live state. That's why an agent can write a note or send you a message while your screen is off. Closing the window doesn't stop the team.
The backend is a stateless Node service (Express plus WebSockets) with Postgres as the source of truth. Here the README got more precise after our first look. The real database schema is the SQL in server/src/db/migrate.ts and its migrations folder. The Drizzle file, schema.ts, is only a partial model. Redis handles pub/sub fan out and presence.
There's also a newer safety net for live updates. When you change a board, document or calendar, Cumora writes the change and a "refresh this" notice in the same Postgres transaction, a pattern called a transactional outbox. If Redis slows down, your screen updates late, but the change itself still lands. Clients catch up by pulling from the API.
Cloud agents each run in their own Kubernetes pod, and a Go driver mounts each agent's workspace on the server. Bring your own agent (BYOA) setups run on whatever machine you choose. Both act through the same Cumora command line protocol, and every model call lands in one cost ledger.
Keeping agents from stepping on each other
This is the part worth slowing down for. When several agents share a room, the server referees. If an agent's reply is stale because new messages arrived while it was thinking, the server holds it and shows it the new messages so it can decide again. Real units of work get atomic claims, so two agents can't grab the same task. A cheaper "small brain" model triages requests first, so the big model isn't woken for noise. The reasoning is written up in docs/COORDINATION.md in the repo.
The project is open source under the MIT license at github.com/yetone/cumora. The desktop app is an Electron shell that updates itself from the separate cumora-releases repository.
Two paths for the agent's brain
This matters if you don't want, or can't afford, to run everything inside one vendor's cloud. Cumora gives you two routes:
- Cumora Cloud. Each agent runs in its own managed pod. Every turn is a multistep tool loop on the OpenAI Responses API, with bash, files, browser, email, memory and skills.
- BYOA (Bring Your Own Agent). You pair your own Mac or server with
npx cumora agent computer, and the agent runs on your local account with one of ten engines: Claude Code, Codex, Grok Build, Cursor Agent, OpenCode, pi, Gemini CLI, Qwen Code, Antigravity or ZCode. The server never sees your provider keys. Details are in docs/BYOA.md.
There's a catch in that list, and it's the most important thing to know before you pick an engine. Only Claude Code and Codex run inside a locked down sandbox by default. That sandbox fails closed: it limits which files the agent can touch, blocks network access from the commands it runs and keeps your credentials away from subprocesses. If the sandbox can't start, the turn stops instead of running with wider access.
The other eight engines only run after you turn on an explicit unsandboxed mode on your machine. That gives the agent the same reach your user account has. It's fine on a throwaway server and a real risk on your work laptop. One more detail from the docs: the Claude Code secure mode works on macOS, Linux and WSL2, but not on native Windows, while Codex's secure mode also runs on native Windows.
If you already pay for Claude Code or Codex, BYOA lets those same tools show up as teammates without Cumora becoming your model provider. If you're still deciding which of the best AI coding agents to standardize on, the sandbox split is now a real factor in that choice.
Who it is for
Small teams that already coordinate in chat and want agents inside that flow. If three to five people run their day in a chat and you want an agent there that remembers what was said and proposes things unprompted, this is the problem Cumora targets.
Developers who already run local coding agents and want them visible as teammates. BYOA is the strongest argument if keeping the brain on your own subscription and your own machine matters to you. For a wider view of the tools in this space, see our AI tools for developers roundup.
People who want agents with initiative without building their own coordinator. The stale reply check, atomic claims and small brain triage are coordination pieces you'd otherwise have to build from scratch. Running it locally also shows you how they work: an empty database is seeded with a starter team of six agents, three humans and nine conversations, with zero messages, so everything you see in chat is generated live.
Who it is not for
Someone who needs a finished, supported product. Cumora says "Free during preview" on its site, and it behaves like one. The waitlist switch on Cumora's public sign up settings read off when we checked on October 7, 2026, and the download buttons are live. The site still carries its old invite only copy in hidden form, though, so access rules could change again.
Someone who needs a full mobile app today. iOS is a TestFlight beta, not an App Store release. Android isn't published at all. You'd have to build it yourself from the android/ folder in the repo.
Someone who doesn't want to manage infrastructure to self host. Running it locally needs Postgres and Redis (Homebrew services work), and OPENAI_API_KEY is the only required setting. Everything else has a default or switches itself off. If running a database and a Redis instance isn't something you want to do, use the desktop or web app instead.
Someone who wants to plug any coding agent in safely. Eight of the ten BYOA engines need the unsandboxed opt in. If your agent of choice isn't Claude Code or Codex, treat the machine it runs on as exposed.
Honest alternatives
Cumora isn't the only project building agent teams, and it isn't the best documented one on every front.
- OpenClaw takes a different route. It's a local gateway that connects channels you already use (WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Matrix, Teams) and configures agents through plain files you can version: SOUL.md for persona, AGENTS.md for instructions and tools, MEMORY.md (which you create by hand), HEARTBEAT.md for what the agent watches on its own, plus hooks. Our OpenClaw guide covers it in depth. The key difference: Cumora is a new team chat app, while OpenClaw plugs agents into the apps you already have.
- Agentic development environments structure agent autonomy differently again. Our Xirp piece is a useful comparison if you're studying how products decide what an agent may do on its own.
- If you're really shopping for a new team chat, our breakdown of Buzz, the Jack Dorsey Slack alternative is the closest neighbor.
The distinction to keep straight: Cumora is a team app with agents built in as members. It isn't a bridge into the chat you already use. If you want agents inside your existing Slack or Discord, look elsewhere.
What it's still missing
Here's where the product actually stands as of October 7, 2026:
- Preview status. It's free during preview, with no published pricing. The download is public today, but there's no promise about what happens after the preview.
- Mobile is partial. iOS is a beta on TestFlight. Android has native code in the repo but no published app.
- Most BYOA engines are unsandboxed. Only Claude Code and Codex get the secure default. Everything else is an explicit trade of safety for compatibility.
- Some features are documented as code, not as a product guide. OAuth login, email through Resend and Cloudflare Email Routing, R2 storage, push notifications, the per user model gateway, invites and metrics are declared in server/src/env.ts, with .env.example annotating a subset. You're reading documented code, not a finished manual.
- No public performance numbers. The benchmarks/ folder holds real model coordination tests (chain, counting, werewolf and kanban scenarios) that the team runs against its own stack. No results have been published, so there are no figures to quote yet. If performance matters to your decision, test it yourself.
If reading this made you want to build agents like these rather than just use them, that's a learnable skill. You don't need to start by running your own servers. You need to understand how agents keep memory, take turns and stay inside safe limits, and a self paced path is a good way to get there.
If you prefer a structured route over a fully self-paced one, a software development bootcamp builds the programming foundations behind agents like Cumora's, through hands-on projects and mentor feedback.