Team chat is the product, not a demo wrapper
Cumora puts AI agents into the ordinary machinery of teamwork. Humans and agents share rooms, direct messages, a roster, Kanban cards, calendars, files, and email. An agent can remember context, claim work, respond to a teammate, or wake because a due task needs attention. That is more useful than another prompt box if your goal is to watch agents coordinate in public and leave behind work the team can inspect.
There are two ways to supply the intelligence. Cumora Cloud runs managed agents through the OpenAI Responses API in per-agent Kubernetes pods. Bring Your Own Agent, or BYOA, pairs a computer or VPS and uses a local Claude Code, Codex, Grok Build, Cursor Agent, or OpenCode CLI under your existing account. The server does not receive those provider credentials. One daemon can host several agents, each with its own working directory, memory, skills, and session.
The React interface spans desktop, mobile, web, and admin shells. The backend uses Express, WebSockets, Postgres, and Redis; email, storage, push, OAuth, and analytics are optional.
Coordination is the most interesting work here
Cumora's strongest idea is that adding more agents creates race conditions, not automatic productivity. Its coordination design documents several safeguards. A freshness check can hold a reply when a newer message arrived after the agent read the room. Atomic work claims reduce duplicate deliverables. A separate duplicate check catches identical replies, while small-model triage filters chatter before an expensive reasoning turn starts. Concurrency limits, pacing, cooldowns, and wake coalescing help several local agents share one provider subscription.
These are practical controls for familiar failures: two agents answer the same request, an old draft lands after the conversation has moved on, or a burst of messages starts too many costly turns. The server can return a held response and ask the agent to reread rather than silently posting stale work. The project also tracks every model call in a cost ledger and enforces that only actual agent turns use the expensive model tier. Those choices show more operational thought than a typical multi-agent demo.
The design still has limits. BYOA isolation is based on separate working directories and per-agent tokens, but agents on one computer share the underlying engine login and global CLI configuration. More importantly, the BYOA guide says engines run with permission prompts disabled. Cumora bounds them to the agent home and server-arbitrated CLI actions, but that is a security decision an operator should review, not a harmless implementation detail.
What happened when we ran it
We cloned commit 12d19ad into a fresh unprivileged Debian container with three CPUs, 6 GB of RAM, Node 22, and no secrets. The repository contained 686 files, about 116,946 lines of source, and occupied 35.1 MB before dependencies.
Installation succeeded in 19 seconds, but it brought in 898 packages and expanded the checkout to 1,053 MB on disk. The build then succeeded in 15 seconds. That proves the code compiles in a clean environment, but the footprint is substantial for an initial evaluation.
The test command failed after 124 seconds. Node's test runner showed 426 passes and 21 failures; its final summary also listed one skipped test and 448 total entries. The last useful error was only ERR_TEST_FAILURE with the message test failed, so the tail does not establish why those cases failed. We will not guess. The reliable conclusion is that the repository's standard test command was not green in our clean sandbox.
The dependency audit was harder to dismiss: 39 known vulnerabilities, comprising 3 critical, 20 high, 14 moderate, and 2 low severity findings. The repository has eight CI workflow files, but no Dockerfile and no top-level tests directory. Tests instead live under server source and worker paths, which matches the package script.
Setup grows quickly after the first build
The README asks for Postgres, Redis, and an OpenAI API key, followed by the setup and combined development commands. Database tables and starter data are created on boot. This is reasonable for contributors, and the environment template clearly separates optional services. It is not the same as a ready-made self-hosted appliance. There is no Dockerfile, and cloud-agent hosting adds Kubernetes, a FUSE workspace driver, storage, email, and push infrastructure as features are enabled.
BYOA has a short pairing command, a doctor check, and supervised service installation for macOS, Linux, and Windows. Windows still has a separate blocker. Issue 58 reports that the published npm package predates a command-resolution fix, causing npm-installed Claude Code or Codex turns to fail with spawn ENOENT. Open issue 75 also says newly installed engines are not detected until the computer is paired again.
v0.4.0 arrived nine days after the repository opened
The public repository was created on August 17, 2026, and its last push was August 27. Release v0.4.0 arrived on August 26 with OpenCode support, attachment path hardening, and dark-theme fixes. GitHub listed 11 open issues and pull requests. Same-day work covered agent context access and local engine detection. The pace is high, but the public history still spans only 10 days.
Documentation is already a strength. The README maps the architecture and repository, while dedicated guides explain BYOA internals, coordination, shipping, email, mobile builds, push notifications, security reporting, and release operations. The interface now ships English and Simplified Chinese, with browser-language detection and English fallback for untranslated strings. Teams needing another locale must add it.
Cumora is worth studying and testing if agent collaboration itself is your problem. Its holds, claims, triage, and shared work surfaces address real coordination failures. For production, the safer answer is wait: a days-old public project with 21 failing tests and serious audit findings has not earned trust merely because its architecture is thoughtful. Run it away from sensitive work, verify the agent permission boundary, and make the test and audit baseline part of any adoption decision.