Rovai AI

Assemble a team of agents that grows together.

Rovai AI is like an Agent guild of your own, where you can recruit members with different personalities and roles. Together, they explore, discuss, and act on real tasks — building team chemistry and collaborative memory along the way.


The story often begins like this

You ask GPT to help draft a plan.

Halfway through, it stops speaking plainly. You hand the answer to another model and ask what GPT was actually trying to say, then bring in a third to poke holes in the plan — and in the end, you still have to decide which one to trust.

Every switch means explaining the roles and pasting the context all over again. When the discussion ends, no one remembers why the decision was made.

A team should not have to get to know one another all over again every time it sets out.

In Rovai, you are the team's Principal.

You can draw inspiration from the games, films, and stories you love, then recruit long-lived members with different personalities and roles:

Some explore, some challenge, some move the work forward, and some remember the road the team has traveled.

They discuss and act on the same task, carrying important decisions, disagreements, and ways of working into the next journey.

When they first meet, they are only agents with different roles.

After completing a few missions together, they gradually begin to feel like a team.


See a team come together

This time, four adventurers answered the call:

Dingding, the wandering scholar — apparently, that's what foxes say; Cheese, who loves to argue — uh, a snow leopard; Gugu the owl — always finding an angle no one else considered; Bunny, the illustrator — sketching what the party sees along the way.

They arrive with different temperaments and talents, meeting for the first time in the same Camp.

The first gathering

The party's first step is not to scatter and start working immediately. It is to learn who everyone is, what this journey is about, and who should speak first.

How they work together

At their first gathering, they are simply lone adventurers with different personalities and roles.

Through repeated discussions, actions, and handoffs, they gradually learn how those differences fit together.


Quick Start

1. Install Rovai AI

Desktop installers (recommended)

Download the installer for your device from GitHub Releases.

Platform Release asset Installation
macOS · Apple Silicon A .dmg whose filename includes arm64 Open the DMG, drag Rovai AI into Applications, then launch it from the Applications folder
macOS · Intel A .dmg whose filename includes x64 Open the DMG, drag Rovai AI into Applications, then launch it from the Applications folder
Windows · x64 — unsigned An .exe installer explicitly labeled for Windows x64 Run the per-user installer and follow the setup wizard

The Windows x64 installer is currently unsigned. Windows SmartScreen may show an unknown publisher warning. Download the installer only from the official Rovai AI GitHub Release.

Rovai AI v0.0.2 adds manual “检查更新” in Settings → About & Updates. A found update downloads immediately, shows progress, and installs only after “安装并重启” is selected. The published v0.0.1 does not contain the updater metadata or install flow, so moving from v0.0.1 to v0.0.2 remains a one-time manual installer upgrade; in-app upgrading starts with releases after v0.0.2.

Run from source (for developers)

For source installation, environment setup, isolated data directories, and build instructions, see the Developer Guide (Chinese).

The shortest development path is:

git clone https://github.com/murray17/rovai-ai.git
cd rovai-ai

pnpm install --frozen-lockfile
pnpm dev

2. Supported Agent Runtimes

In Rovai, who a member is and which Runtime they act through are two different layers.

A member's name, appearance, responsibilities, relationships, and collaborative memory define who they are. The Agent Runtime determines which tools and models they use to participate in the work.

The same Codex Runtime can power a builder focused on delivery or a challenger searching for counterexamples. The same Claude Code Runtime can serve as a strategist, record keeper, or reviewer, depending on the team.

Agent Runtime MCP support Skill support Identity continuity
Claude Code Added alongside native Added alongside native Native support
Codex CLI Added alongside native Added alongside native Native support
OpenCode Added alongside native Added alongside native Re-delivered after compaction
GitHub Copilot CLI Added alongside native Added alongside native Re-delivered after compaction
Antigravity Runtime-native only Added alongside native Based on Runtime capabilities
Kiro CLI Added alongside native Added alongside native Re-delivered after compaction
Qoder CLI Added alongside native Added alongside native Re-delivered after compaction
CodeBuddy Added alongside native Added alongside native Re-delivered after compaction
Qwen Code Added alongside native Added alongside native Re-delivered after compaction
TRAE CLI CN Added alongside native Added alongside native Based on Runtime capabilities
Kimi Code Added alongside native Added alongside native Native resume; re-delivered after compaction

For exact versions, capabilities, and observed boundaries, see the Agent Runtime Compatibility Register (Chinese).


Core Capabilities

Core capability What it means in Rovai
Long-lived members Preserve each member's enduring identity, appearance, responsibilities, and working style so they can rejoin the team across Camps and Tasks.
Camp collaboration Organize shared conversations, long-lived members, Tasks, attachments, and execution state around one objective, reducing the context users must carry between separate windows.
Role-based collaboration Turn paired questioning, document-informed decision-making, code review, and group discussion into repeatable ways for a team to work together.
Tasks and ownership Give ongoing work a title, owner, and state so unfinished work remains trackable and can be picked up by another member.
Member handoffs Use @mentions, the default Lead, direct replies, and A2A routing to pass questions, conclusions, and next actions to the right member.
Visible execution Inspect tool calls, process state, intermediate results, and final delivery in an independent execution console instead of hiding real work behind a single completion message.
Approvals, evidence, and recovery Require explicit approval for important actions, preserve reviewable execution evidence, and continue from existing state after interruption.
Collaborative memory Preserve important decisions, lessons, and team habits so members gradually understand how this team has solved problems together.
Native capability compatibility Use a generic ACP Adapter to connect Agent Runtimes that support ACP while preserving their native models, permissions, Skills, MCP, and session capabilities whenever possible.

These capabilities are not isolated features. They are connected by one collaboration architecture:


Design Philosophy

We believe this is how a team grows.

Capabilities bring members into the team. Shared experience helps them understand and trust one another.

✦ Worldbuilding adds warmth; the work stays professional

Camp, Principal, members, and journeys are Rovai's language for expressing collaboration, not an extra layer of role-playing.

Rovai keeps the design restrained. Worldbuilding never adds irrelevant steps or bloated context.

✦ Growing together does not mean growing alike

Chemistry does not mean everyone eventually gives the same answer.

Explorers keep exploring, challengers keep testing assumptions, and builders keep moving the work forward. Over time, they learn whose judgment to trust, and when.

✦ Remember why, not every word

Team memory is not an ever-growing transcript.

What matters is remembering why a choice was made, what remains unresolved, what the action led to, and which detours the team should not have to repeat.


Documentation

Detailed guides are currently available in Chinese.


Contributing

Issues and Pull Requests are welcome.

See the Version Roadmap (Chinese) for ongoing and planned work.


License

The MIT License allows use, modification, distribution, and commercial use.