Cumora puts AI agents in a chat room next to humans, and the part of its README worth reading closely is how those agents avoid colliding. The server-side mechanisms it describes read like a database's answer to concurrent writers.
Cumora is cross-platform team chat where agents sit on the same roster as humans, with the same DMs, group conversations, Kanban board, and calendar. The README is explicit that agents "don't just answer when poked." They hold personas and memory, claim work, and send and receive real email. The README describes several mechanisms for containing collisions between them.
Stale replies get held
The coordination layer rests on three mechanisms the README names.
The first is a seen-cursor freshness gate. If newer messages arrived in the room after the point an agent had seen, its reply is HELD, and the agent is shown the newer messages so it can decide again.
The second is atomic claims on real units of work. The server arbitrates who holds each claim.
The third is a small-brain triage gate that shields the big model. The config reflects the split directly: OPENAI_MODEL and OPENAI_MODEL_SUPPORT map to big-brain and support-brain models, and a CI guard, npm run guard:big-brain, enforces the rule that only agent turns may use the big model. The README calls it a CI guard.
Design notes live in docs/COORDINATION.md, which the README describes as covering defense layers and anti-patterns. The repo also carries a benchmarks/ directory of real-LLM multi-agent coordination benchmarks, with scenarios named chain, counting, werewolf, and kanban.
The same discipline in the backend
The concurrency mindset is not limited to agents. The backend is a stateless Node service on Express and ws, with Postgres as the source of truth and Redis for pub/sub fan-out and presence. Durable board, document, and calendar writes enqueue realtime invalidations in a transactional PostgreSQL outbox. Any number of instances can drain that outbox through leased SKIP LOCKED claims.
The README spells out the behavior during Redis degradation: Redis degradation delays live refresh but never changes the command result, and clients reconcile by pulling the API. Live updates ride on top of a correct write rather than standing in for it.
Spend gets a single place to land too. Every LLM call, whether from a cloud agent or a local one, is recorded in one llm_calls cost ledger.
Two brains, one protocol
Agents run on one of two paths. On Cumora Cloud, each agent gets a managed per-agent Kubernetes pod, orchestrated with kubectl from the server, with a Go FUSE driver mounting its server-side workspace. Turns run a multi-hop tool-calling loop on the OpenAI Responses API, with tools covering bash, files, browser, email, memory, and skills.
The other path is BYOA, Bring Your Own Agent. You pair a Mac or VPS with npx cumora agent computer and the agent runs on your local provider account. The README lists Claude Code, Codex, Grok Build, Cursor Agent, OpenCode, pi, Gemini CLI, Qwen Code, Antigravity, and ZCode. Claude Code and Codex run inside fail-closed filesystem, command-network, and subprocess-credential boundaries by default. The other engines need an explicit unsandboxed compatibility opt-in. The server never sees your provider keys.
Both paths act on the world through the same cumora CLI protocol.
Running it yourself
Local setup needs Postgres, Redis, and an OPENAI_API_KEY, which is the only hard-required variable. Run npm run setup, then npm run dev:all to start the Vite renderer on port 5180 and the API server on 5181, with migrations applied automatically. An empty database seeds a starter team of 6 agents, 3 humans, and 9 conversations with zero messages, so everything that shows up in chat is produced live.
One warning from the README deserves attention before you trust a green run. Without INTEGRATION_DATABASE_URL, the integration suite prints [integration] skipped and exits 0, which looks like a pass. When it does run, it truncates every table, so point it at a throwaway database.
GitHub: https://github.com/yetone/cumora
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)