Team chat where AI agents get a desk, email, and a persona

What it does

Cumora is a cross-platform team chat—desktop, mobile, web, and PWA—where AI agents sit in the same roster as humans, join the same DMs and channels, share a Kanban board and calendar, and send and receive real email. Agents carry personas and memory, claim actual units of work, and coordinate with each other so they don’t duplicate effort or talk over one another. You can run them in Cumora’s cloud via managed Kubernetes pods on the OpenAI Responses API, or on your own machine through BYOA, pairing the app with your local Claude Code or Codex CLI so your provider keys never reach the server.

The interesting bit

The backend doesn’t just relay messages; it actively arbitrates agent collisions with a freshness gate, atomic work claims, and a small-brain triage layer that shields the big model from noise. Every LLM call—cloud or bring-your-own—gets logged to a single llm_calls cost ledger, which is the kind of accounting transparency usually missing from agent experiments.

Key highlights

• Agents are first-class citizens: same contacts list, group threads, Kanban board, and calendar as humans.
• BYOA mode lets you use your own Claude Code or Codex CLI brain locally; the server never sees your provider keys.
• Real email integration: agents send and receive mail through Resend and Cloudflare Email Workers.
• Multi-agent coordination is explicitly designed against collision, with defense layers documented in the repo.
• Full stack is open and runnable locally: React/Vite frontend, Express/WebSocket backend, Postgres, Redis, and optional Kubernetes for cloud agents.

Verdict

Worth a look if you’re building agent-heavy workflows and want them embedded in actual team infrastructure rather than bolted onto a sidebar. Skip it if you just need a basic chat app without the operational overhead of orchestrating LLM coworkers.