Local agents stay local and identity never leaves the machine, while multi-layer virtualization brings Your Claude Code and your teammate's Codex into one shared cloud Room, working in parallel in real time.

HONG KONG, Aug. 28, 2026 /PRNewswire/ -- The Tutti team today opened Early Access for Tutti • VM, the industry's first multi-user, multi-agent, real-time collaboration space and the first product to combine the best of local tools and cloud tools in one. In a Tutti • VM Room, two people's coding agents can edit the same file at the same time without colliding. One agent can pick up another's work with no re-briefing. A website still running on someone's localhost can be shared with a single link. And through it all, every agent keeps running on its owner's computer, on the subscription its owner already pays for.

Agents Got Faster. Organizations Didn't.

Mainstream coding agents such as Claude Code and Codex run in isolated local environments. An agent takes a task, makes decisions, runs commands, and hands back a result, and only its own user sees the process. The collaboration itself runs through people. The agent summarizes to its user, the user explains to a teammate, the teammate briefs their own agent, and every hop loses information.

Unless the decisions, the file changes, and the reasoning behind them are relayed word for word, the next agent has to re-understand the project from scratch and guess what another agent was thinking from a few lines of briefing.

Most multi-agent products today have agents talk through group-chat messages or hand off sequentially under an orchestration layer. Both pass along compressed summaries. The real work stays invisible, conflicts surface only after the fact, and the more agents there are, the more rework follows. Many teams share the same experience: agents deliver noticeably faster, yet the organization as a whole does not, because the time goes into summarizing and handing off.

Three Problems Nobody Had Solved

For everyone to keep using their own machines and local tools while working in parallel in one persistent cloud workspace, three problems had to be solved first.

The strengths of local and cloud tools are mutually exclusive by design. Local-runtime tools operate directly on real projects and local files, but the machines are isolated from one another. Cloud-runtime tools are built for sharing, but they cannot use local tools, accounts, configurations, or files, and they require users to hand their identity to the platform.

Agents on different computers have no real-time channel. Context and outputs travel through slow uploads and downloads. By the time tens of megabytes of code or hundreds of megabytes of video finish syncing, parallel work has already been interrupted.

Parallel work also requires solving collaboration conflicts at the root. The ways agent commands and file operations can collide are impossible to enumerate: two collaborators installing npm dependencies at the same time, or one agent writing to a file while another moves it away. These cases go far beyond what the application-layer algorithms behind online documents can cover.

How It Works

Tutti • VM separates identity, execution, and sharing, places each where it belongs, and merges them back into one coherent agent environment.

Identity stays on the user's computer. Claude Code and Codex keep running as local tools on the user's physical machine. Login sessions, subscriptions, credentials, API keys, SSH keys, and corporate SSO never leave the machine, and the cloud never holds them. Agent requests still go out from the user's real device and network identity, so model providers' risk controls, enterprise IP allowlists, and internal services keep working.

Execution happens in a local virtualization layer. Multi-layer virtualization maps the agent's runtime into a managed Linux environment, where the agent's process calls and file operations are intercepted, parsed, and redirected at the kernel level. Weighing correctness, performance, and sandbox cost, the system decides whether each instruction ultimately runs on the physical machine, in the VM, or in the cloud.

Sharing and collaboration happen in the cloud Room. Only instructions that affect collaboration and outputs reach the cloud. Whatever an agent instruction issued through the Room produces, the process and the results alike, is born in the Room. There are no second copy, so even large context and outputs never travel through upload and download. To users, everyone's live working state plays out in the Room, visible the moment it happens and ready to be built on.

On top of these three layers, Tutti • VM built its own real-time collaboration technology for agents, currently the only one of its kind in the industry, which pushes conflict handling down to the filesystem layer. The filesystem is the one language every tool speaks. Claude Code, Codex, vim, or a ten-year-old shell script will not change a line of their own code for collaboration, but they all read and write files. Built into the filesystem, collaboration becomes transparent to every tool, and thirty years of existing software "learned" to collaborate overnight. The work spans six engineering domains, from OS kernels and multi-layer virtualization to distributed consistency, collaborative editing, network and security engineering, and local-to-cloud performance optimization. Because only lightweight collaboration instructions reach the cloud, per-user sandbox cost comes to 20 to 25 percent of what a typical cloud solution costs.

Tan Zeng, Founder of Tutti said: "Think of the movie Ready Player One. Every player stays in their own separate physical space, and once the headset goes on, their avatars connect in one virtual world. Tutti • VM is the device that brings agents into that world. Local agents don't change at all. We connect everyone's previously isolated agents into one Room, and real-time parallel collaboration and context handoffs grow out of that."

Inside a Room

A live shared workspace and outputs. Everyone's local agents live in one cloud workspace, where environments, services, and dependencies are configured once. What each person is discussing with their agent, which files are being changed, and what is already done stay visible to the whole Room. Websites, videos, and documents made by agents are shared as they take shape. When a localhost site changes, collaborators see it immediately. When an 800MB video finish rendering, collaborators can watch it in full right away.

Parallel work. Different people's local agents can build the same website or the same document at the same time, even editing the same file at the same time. Tutti • VM avoids and resolves conflicts at the filesystem layer, in the background, and there is no merge to untangle afterward.

Context handoffs and borrowed agents. Typing @ references any collaborator's session, file, or output, and another agent continues from the current progress with no re-briefing. A collaborator can also authorize their agent for someone else to use. The borrowed agent picks up the borrower's own context, so it works like the borrower's own agent. Authorization can be revoked at any time, and the owner's login never leaves their machine.

Built-in collaboration apps. Tutti • VM ships with group chat, docs, slides, image and video generation, vibe design, task management, and an agent board. Their outputs stay in the Room as well, ready to be referenced by people and agents alike.

Availability and What's Next

Tutti • VM is free during Early Access and follows a bring-your-own-subscription model, currently supporting Claude Code and Codex, so users keep the agent subscriptions they already pay for. Seat-based plans will follow. Rooms are invite-only. Invite codes are limited and released in batches, and one code is enough to create a Room and invite collaborators. Codes are available at tutti.sh.

Tutti • VM will also offer an SDK and private enterprise deployment, giving a team's local agents real-time collaboration, unified management of context and outputs, and organization-wide agent sharing. Updates will follow.