Tutti

You will feel this on the second agent. The trait that decides it in daily use is that conversations, files, running tasks and app outputs live in one shared state, so asking a second agent to build on what the first produced is a reference rather than a copy-paste ritual. Goals decompose into subtasks in the same place, which keeps the thread of what is happening in one window.

Pick it when coordination between agents is the actual work. Pick Agent Teams AI if you want that coordination expressed as a board, or a single terminal agent if you have never wanted a second one.

Shared state removes the step where a person reads something. When agents reference each other's outputs directly, an early misunderstanding propagates as fact rather than as a suggestion someone chose to accept, and the single approvals surface is exactly the interface that trains people to approve quickly. Decomposition makes it worse, since the subtasks a human never specified are the ones nobody checks.

What it does right: the agents run locally on subscriptions you already hold, so nothing is resold to you and no new party sits between you and the model.

Two points about the published material. 1. The design uses a remote connector as an HTTP client and a local aggregated connector as a server exposed to the hosted agents, which is a sound arrangement: tools are collected once and presented uniformly rather than configured per agent. 2. That arrangement is described in architecture notes inside the repository, so a reader who stops at the front page will not know it exists.

Nothing is claimed on any benchmark. The documentation is thin in the ordinary way of very new projects, and thinnest exactly where the reusable idea is.

The structure is legible: give away the single-user desktop edition, gate the multi-user hosted one behind invitations, and charge later. Invitation gating during early access is a supply story that doubles as a demand signal, and it works only while the waiting is worth something. Nothing is priced yet, so pricing power is entirely hypothetical.

Three and a half thousand stars against a June 2026 first release is a fast start and a thin base. Likely acquirer: a collaboration vendor wanting multi-agent workspace mechanics. Position: adopt the free edition, and do not plan around a paid tier that does not have a number yet.

Licensing is nothing and the gaps are structural. Supported platforms are macOS and Windows, which leaves the engineers on Linux workstations without a path, and there is no headless mode, so this stays a desktop habit and never becomes a step anything downstream depends on. Identity is absent: no single sign-on, no SCIM, no audit trail describing which agent did what under whose name.

The multi-user edition is where those questions would matter most, and it is invite-gated, so I cannot evaluate it. Onboarding is half a day. Not yet, and ask me again when Rooms are open.

Apache-2.0 with no additional clauses, which makes the fork question boring in the best way. It consumes MCP servers as a client and exposes an aggregated one back, so my existing tools are available to every agent it hosts without me configuring them three times. The agents run on my machine using logins I already have.

The limit is inference: the model never runs on my hardware, so this is a coordinator for other people's clients rather than something I can take entirely offline. Good licence, real protocol support, and one door I cannot open.