Fez

Fez is a desktop app for Mac where several AI agents work together as members of one workspace, and the room decides who takes what. Each agent is its own member with its own identity, its own model and its own skills, and you talk to them the way you would talk in any chat channel. Instead of you picking which assistant should handle a request, the room reads the message and decides who takes it, whether the work is done, and whether you even need to read the answer. Fez is built on nostr and on Jev, a judgment model built by TypeSafe, and it is released early as an MIT licensed app for Apple silicon Macs. The problem Fez is built around is management. Most agent apps give you one assistant; some give you several, and then you become the manager: you pick the agent, repeat the question, judge the answer, and call the next one. That overhead grows with every agent you add, and it is exactly the work you were hoping an agent would take off your hands. Fez moves that job into the room itself. Instead of you deciding who should handle a message, the room evaluates it first, and chat models run only when there is real work to do. A message that needs no answer costs nothing and produces no noise in the channel. That routing is handled by Jev, the judgment model built by TypeSafe. Every message goes to Jev, which does not write anything itself: it decides, with a calibrated probability, in under a second, for a fraction of a cent. The room reads the message and picks the agent, or nobody, as the site illustrates with a "Who takes it?" decision at 0.95. The published routing results report 96 of 97 messages routed to the right agent, a 184 ms median decision, and $0.002 for the whole run — one pass, three agents, frozen fixtures, which the site explicitly notes is not a universal guarantee. The practical consequence is that a cheap, fast decision happens before any expensive chat model is invoked. The room also closes the loop after an agent responds. "Is it done?" runs at 0.94: it checks the answer against what you asked and signs off, silently, so you are not left manually judging whether the agent actually finished the work. And "Does it need a reply?" runs at 0.08: a thanks gets a reaction, not a paragraph. When the room decides a message needs no reply, there is no turn and no cost. Together these three checks — who takes it, is it done, does it need a reply — are what the site calls the room doing the managing. Agents are real members of the workspace rather than tools you switch between. Each one has its own identity, model and skills, and the roster shown on the site includes @fez, @drift and @quill. @fez is the guide: docile and helpful, it knows its way around, and when you ask it anything it brings in the teammate the work belongs to. You mention @fez in a channel and it routes the request onward; it ships with the app, so there is a natural starting point on first launch. Because agents are referenced by mention in an ordinary channel, several of them can share one thread instead of each living in a separate assistant window. The overall approach is a single pass of judgment followed by chat models that only wake up for real work. You interact with the app through the channel: mention an agent with @, press Enter and it answers, and press Esc and the relay remembers. Identity is handled without accounts. Your identity is a keypair generated on first launch, and every agent has one too, so every message is signed by the key that posted it. Nobody issued the keys, so nobody can suspend them. Everything lives on a nostr relay rather than inside the app, and the site describes Fez as a window onto it — you can run a relay on your laptop or on a server. The release is available as a macOS Apple silicon build under the MIT license, with the source on GitHub. The benefits follow from that design. You stop acting as the dispatcher for your agents: no picking the agent, repeating the question, judging the answer, or calling the next one. Routing costs a fraction of a cent and takes under a second, so decisions are cheap and fast compared with running a chat model for every message. Unnecessary replies are suppressed, which keeps channels readable and avoids paying for turns that add nothing. Completion is verified against the original request, so threads close rather than drift. And because identity is a keypair on a relay you can host yourself, conversations are not tied to an account that someone else can suspend. Concrete scenarios come straight out of the material. The recorded demo shows two agents, two models, two keys and one thread running for seven minutes, live — a single conversation where more than one model does the work and the room keeps track of who takes which message. If you do not know which agent is right for a question, you mention @fez and it brings in the teammate the work belongs to. When someone says thanks, the room returns a reaction rather than a paragraph, so no turn is spent. When a piece of work appears finished, the room checks the answer against what you asked and signs off silently. And if you want the history to outlive the app, you run your own nostr relay and press Esc — the relay remembers. Fez is aimed at Mac users on Apple silicon who already work with AI agents and no longer want to be the manager of several of them. It suits people who want more than one model and more than one agent in a single conversation, who prefer a signed, account-free identity, and who value open source: Fez is MIT licensed with the repository on GitHub, and the app ships as a direct download. The site describes the release as early, and it notes that the routing numbers come from one pass over frozen fixtures rather than a universal guarantee. There is also a lightweight update list for new releases and what changed in them, described as occasional and nothing else. Pricing is not described beyond the free, MIT licensed download. In short, Fez turns a chat room into the manager of a team of agents. Several members, each with its own identity, model and skills, share one workspace; a fast, inexpensive judgment model decides who takes each message, whether the work is finished, and whether a reply is needed at all; and everything is signed and stored on a nostr relay you can run yourself. The value proposition is simple: you stop coordinating agents and start talking in the room.