Rovai AI
Desktop and browser workspace that keeps coding-agent teams together across sessions, tasks and projects.
- Source repo
- murray17/rovai-ai
- Stars
- ★ 276
- Last updated
- 1d ago
- License
- MIT
- Primary language
- Rust
- FA score
- 77/100 · Good
At a glance
- Works with
- Universal · cross-platformCodex · Claude CodeChatGPT (Partial support)
- You'll need
- Typical use
- You already have Claude Code or Codex CLI installed and signed in, and want an implementer and a reviewer to work the same project with different Agents, models and permissions while handing work back and forth.
- Main limitation
- The software is free but you must supply and sign in to your own coding Agents and model services, so real running costs depend on those providers.
- Website
- rovai.dev ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v0.5.14 (2026-11-17)+1.7 ★/day over 7 days
- Source review
- 77/100 · Good
What does this agent do, and when should you use it?
Rovai AI is a workspace for long-lived coding-agent teams, shipped as a desktop app for macOS (Apple Silicon and Intel) and Windows x64, plus browser access and a standalone Rovai Server you can host yourself. It brings already-installed, already-signed-in coding Agents — Codex CLI, Claude Code, Pi Coding Agent, DeepSeek Harness, OpenCode — into shared conversations in which each teammate has a name, role, responsibilities and working principles, with its Agent, model and permissions chosen separately. Rovai itself does not perform the coding work: the coding Agents act with their own tools, authentication and configured model services, so model choices, permissions, Skills and MCP support depend on the Agent and host platform. Execution stays visible through execution cards where you follow tool activity, answer approvals, preview files and inspect changes, while tasks record responsibility and execution records show what actually ran. Longer goals live on the Mission board with scheduled work and collaborative memory, and finishing a run, completing a mission and merging code are kept as three separate actions. Desktop and standalone Server share the Host implementation but keep separate data, and channel integration for Feishu and DingTalk is provided by Desktop.
Rovai reads and maintains the team's collaboration state: teammate profiles (name, role, responsibilities, working principles), shared conversations and focused one-on-one threads, Mission board goals with project and team, task ownership, scheduled work and saved collaborative memory. It invokes the coding Agents you have installed and signed in on the host; those Agents run with their own tools, authentication and configured model services, which is why available models, permissions, Skills and MCP capabilities follow the chosen Agent and host platform. Execution surfaces as cards where you follow tool activity, respond to approvals, preview files and inspect changes; tasks record responsibility while execution records show what ran, and finishing a run, completing a mission and merging code are three distinct actions. For remote work, the desktop app can enable Web access or you can run an independent Rovai Server on your own host and continue in a browser, with Agents and project files staying on the host you connect to; the desktop app also connects to Feishu and DingTalk to send requests and receive results. Building from source uses git clone, pnpm install --frozen-lockfile and pnpm dev, and the engineering guides are primarily in Chinese.
- You already have Claude Code or Codex CLI installed and signed in, and want an implementer and a reviewer to work the same project with different Agents, models and permissions while handing work back and forth.
- You are running a multi-day goal such as reworking a download flow or preparing a release, and need a Mission board that keeps the goal, project, team and accumulated file changes in one place.
- You want durable conventions — interface rules, style guides, release steps — stored as collaborative memory so the next teammate does not start from zero.
- You want to leave the desk, open the workspace in a mobile browser, and follow execution while code and Agents stay on your own host.
- You prefer to file requests and receive results inside Feishu or DingTalk and want your Desktop teammates reachable there.
- You want to wire existing services into the team's workflow through native Skills or MCP without switching the coding Agent you already use.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pnpm install --frozen-lockfile
- Runtime
- Rust 1.85+ · Node.js 24+
- Install notes cover
- Windows · macOS
How do you install or deploy this agent?
Desktop (the documented primary path): download the build for your platform. The README lists macOS Apple Silicon, macOS Intel and Windows x64 targets.
https://rovai.dev/download/Self-hosted server: install Rovai Server on your own host. Server packages, host requirements, release limitations and setup are documented separately.
https://rovai.dev/docs/server-install.htmlBuilding from source requires Rust 1.85+, Node.js 24+ and pnpm.
git clone https://github.com/murray17/rovai-ai.git
cd rovai-ai
pnpm install --frozen-lockfile
pnpm devBefore you start, install and sign in to at least one supported coding Agent on the host (Codex CLI, Claude Code, Pi Coding Agent, DeepSeek Harness, OpenCode and others), then confirm it in Settings → Agents.
How do you use this agent?
The README quick start is three steps:
- Prepare one Agent: install and sign in to a supported coding Agent on the host, then check it in Settings → Agents.
- Choose a teammate: set their Agent, model and permissions. One teammate is enough to get started.
- Give it a small task: open a project conversation, ask the teammate to explain the project, and inspect the execution before assigning a change.
For a larger goal, create it on the Mission board with its project and team, continue in that conversation, and inspect the accumulated file changes and delivery. To continue from another device, enable Web access in Rovai Desktop or run a standalone Rovai Server and open the workspace in a browser; the desktop app can also connect to Feishu and DingTalk.
What are this agent's strengths and limitations?
- Teammate identity outlives individual sessions: name, role, responsibilities and working principles persist, and each teammate picks its own Agent, model and permissions.
- Agent connectivity is not tied to one vendor — the README names Codex CLI, Claude Code, Pi Coding Agent, DeepSeek Harness and OpenCode, with models and capabilities following the chosen Agent.
- Execution is visible and separated from ownership: follow tool activity, answer approvals, preview files and inspect changes, while tasks record responsibility and execution records show what ran.
- Deployment boundary is explicit: Desktop and standalone Server share the Host implementation but hold separate data, and a browser connection does not migrate or sync it; Agents and project files stay on the host you connect to.
- MIT licensed, free to use, modify, distribute and use commercially.
- The software is free but you must supply and sign in to your own coding Agents and model services, so real running costs depend on those providers.
- Capabilities are bounded by the chosen Agent and host platform: model choice, permissions, Skills and MCP support all depend on them, and the README points to a separate compatibility document.
- Engineering documentation — developer guide, architecture and runtime-compatibility evidence — is primarily in Chinese, which raises the barrier for English-only readers.
- Remote access is not synchronization: Desktop and standalone Server each keep their own data, so teams must decide up front which host holds the work.
- There is a hard prerequisite: at least one supported coding Agent must be installed and signed in on the host, so a fully offline or Agent-less environment cannot run it.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Rovai AI This agent | 77 · Good | — | ★ 276 | 1d ago | Rust | Codex · Claude Code |
| Codync | 74 · Some gaps | Desktop appFree | ★ 2.3k | 1d ago | Rust | ChatGPT · Codex · Claude Code · Claude.ai |
| Novadeck | 55 · Major gaps | Desktop appFree | ★ 55 | today | TypeScript | Codex · Claude Code |
| tlbx | 73 · Some gaps | Self-hosted serviceFree + model costs | ★ 119 | today | C# | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The Host CLI uses an explicit private data directory, reads the Web management token from stdin, and requires an explicit flag for insecure LAN HTTP. Installation notes say upstream credentials are not managed; the security policy asks reporters to redact sensitive data, and postmortems disclose an MCP isolation failure and its remediation. These support data-flow and risk traceability. Runtime permissions depend on the Agent and host, and the sample does not fully show permission and revocation flows, so these scores are not higher. MIT and repository author metadata are visible, but the publisher is unverified and maintenance ownership is unclear.
The repository lists CI, type-check, compile, and test entry points; two postmortems document defects, boundaries, and uncovered cases. However, they note validations that were not rerun or had limited platform coverage, and one describes a historical architecture transition, so reliability evidence is moderate. Dependencies use locked installation and stated version requirements, but the sample cannot establish availability or failure handling for every critical dependency. Setup status tables and incident messages give actionable guidance, but not every failure path is shown.
The documentation targets lasting coding-agent teams and covers desktop, standalone server, browser access, tasks, collaboration, and memory. Skill descriptions state trigger and exclusion boundaries, and the product says model, permission, and MCP capabilities depend on the Agent and platform. The sample does not fully enumerate product capability limits. Installation guidance covers macOS, Windows, and source development, and explicitly names some unqualified platforms; environment fit is supported but not comprehensive.
Documentation maps, version authority routes, and architecture guidance are clear. Installation steps, FAQs, examples, platform limits, and the MIT license are concrete. Types and APIs retain compatibility naming such as Thread/Camp; version 0.5.9, version governance, and incident records provide some change history, but the sample lacks a complete user-facing changelog. Contribution and security-report paths exist, while a postmortem says action owners must still be mapped to maintainers, leaving responsibility only partly clear.
The product combines multi-agent collaboration, execution visibility, file-change inspection, and continuing memory in one workspace. The README gives a usable starting path, and contract types expose runtime configuration and capability data. These provide concrete evidence of utility. Running agents, configuring runtimes, and enabling remote access carry setup and operating costs; the sample contains no cost-benefit measurement, so this criterion is not full marks.
README claims can be compared with contract types, Host arguments, package metadata, and CI workflows. Postmortems identify concrete defects, remediation versions, validation scope, and remaining gaps, while separating observations from inferences. The packet is only a source sample, and postmortem validations were not independently run or checked in this review, limiting cross-source confirmation.
- This is a static review with low confidence; no code, tests, or runtime behavior were executed.
- The product connects to external Agent runtimes, model services, MCP, and messaging channels. Configure permissions and remote access as documented, and keep tokens, credentials, and user data out of logs, screenshots, and reports.
- Postmortems record a Codex MCP isolation incident and DSH model-configuration and response-order defects. Although the documents describe fixes, those records do not establish that the target revision's implementation was independently verified.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.5.14 Rovai AI v0.5.14
Release notes excerpt
<!-- lang:zh-CN --> 本次更新新增 Goose 智能体,并调整执行卡片和消息多选入口。 - 【功能】新增 Goose 智能体。 安装并配置官方 Goose 后,可在智能体设置中检查,并在队员管理或首次引导中选用。支持 Windows x64,macOS(Apple 芯片和 Intel)与 Linux x64 提供预览。 - 【功能】Goose 支持会话恢复和执行中插话。 重新打开会话后可以接着之前的工作继续,任务运行时也可以补充指令。 - 【功能】Goose 接入 Skills、MCP 和工具审批。 可以在 Skills 页查看本机 Goose 发现的技能,在会话中使用分配给队员的 MCP 工具,并处理工具审批。 - 【功能】可以在会话中查看 Goose 的文件变更。 新建文件可查看新增内容,编辑文件可查看具体修改片段。 - 【交互】Goose 模型列表支持手动刷新。 本机 Goose 的服务商或模型有变化时,可以在模型选择器中刷新。个别服务商刷新失败时,成功刷新的结果仍会显示。 - 【修复】Goose 的思考强度随所选模型更新。 队员设置和首次引导都会读取当前模型支持的档位,也可选择「跟随模型默认值」。 …View on GitHub ↗ -
v0.5.13 Rovai AI v0.5.13
Release notes excerpt
<!-- lang:zh-CN --> 本次更新新增长期记忆开关,支持在会话里直接下载队员发来的文件,并改善长会话的输入、翻阅和恢复。 感谢 @NorethSea 在 Issue #1074 中提出关闭 Rovai 内置长期记忆的建议,也感谢 @liuqiong0319 反馈 GitHub Copilot 无法运行的问题。 - 【功能】新增长期记忆总开关。 默认开启,可在「通用设置 → 会话」下切换「记忆」,也可以在记忆页操作。同一实例的桌面端、网页端和移动端共用设置。关闭后暂停队员读写和记忆页管理,已有记忆和待审核项会保留,重新开启后可继续使用。 - 【功能】可直接下载队员发来的文件。 鼠标悬停在文件行或用键盘聚焦时,右侧会显示下载按钮。触屏直接显示该按钮,桌面端可以选择保存位置。 - 【性能】改善长会话的输入和滚动。 公屏只渲染阅读位置附近的消息,队员的实时输出只更新对应执行。向上加载历史时保留阅读位置。关闭搜索后,页面以用户最新的滚动、发送或定位操作为准。 - 【界面】执行卡片用一个数字显示已统计的 Token。 执行中和结束后都可查看本次执行已记录的用量,例如 12.5K。点击可查看输入、输出、缓存和耗时明细。 …View on GitHub ↗ -
v0.5.12 Rovai AI v0.5.12
Release notes excerpt
<!-- lang:zh-CN --> 本次更新新增会话分叉和 Kilo 智能体接入,OpenCode 也可以在执行中插话。 感谢 @arschlochnop 在 Issue #1042 中反馈长会话难以分段继续的问题,也感谢 @caoshuisi 在 Issue #997 中提供本机 CLI 升级后版本信息滞后的复现。 - 【功能】支持会话分叉。 可从侧栏会话菜单复制全部公屏历史,也可在队员消息的三点菜单选择「从此处分叉」,只复制到该条消息。新会话沿用原项目目录和在队成员,保留消息引用和附件入口。 - 【功能】接入 Kilo 智能体。 可沿用本机 Kilo CLI 的登录和模型配置,在 Rovai 中选择模型与推理强度、使用 MCP,并查看文件变更和子代理进度。Windows x64 可用,macOS(Apple 芯片和 Intel)与 Linux x64 以预览方式开放。 - 【功能】OpenCode 支持执行中插话。 使用 2.2.7 起的稳定 2.x 版本时,可以在队员执行中补充指令,后续执行也可以接续原会话。 …View on GitHub ↗
FAQ
Does Rovai cost anything?
Does Rovai itself write or run the code?
If I open the workspace in a browser, is my desktop data synchronized?
Which operating systems can run the desktop app?
When is work actually finished?
Is Rovai AI open source?
Is Rovai AI safe to use?
Is Rovai AI still maintained?
Who makes Rovai AI?
What is Rovai AI's official website?
Thanks — we use this to improve the page.