Viden

The deciding trait is that everything is in one place. The conversation, the pending approvals, what the workspace currently looks like, which tasks are live, the diagnostics and whether the provider is even healthy, all on one surface. Anyone who has hunted for the reason an agent went quiet, only to find an outage, will appreciate that last one immediately.

It is a lot of interface for a young project. Pick it if you like watching the machine work. Pick something simpler if you would rather not.

A permission gate works by seeing an action before it happens. A pseudo-terminal lane delivers keystrokes into a live shell, where the unit of action is a character, so the checkpoint that governs everything else cannot govern that. The row also carries an experimental surface without saying what experimental means for the guarantees around it.

There is no container underneath any of this. What it does right is admitting which surface is unfinished, which is more than most projects of this size manage.

1. A single chokepoint for file writes, shell invocations, version control and test commands is the architecture that makes a permission model auditable, because there is one place to read and one place to change. The common alternative, a check duplicated at each call site, is where omissions live. 2. Reading and writing sharing a path also means logging is uniform by construction.

2. No evaluation of the loop is published, which is expected. The architectural claim is the contribution and it is a real one.

The product is betting on a future where developers supervise several agents at once, which is plausible and not yet true for most people, so the payoff is deferred while the maintenance is immediate. Under a hundred stars means nobody is sharing that bet, and there is no entity, revenue or service behind it.

Moat: none; a supervision surface is a feature the agent vendors will absorb. Likely acquirer: none. Likely path: ahead of its market and out of energy before the market arrives. Position: watch, do not depend.

Local storage by default is the right posture and it creates the discovery problem. Every record of what an agent did sits on the engineer's own disk, so producing evidence for an auditor means collecting it desk by desk, and retention is whatever each person's machine happens to do.

There is no identity integration, no directory sync, no export and no unattended mode, so it stays a personal cockpit rather than a stage I can gate a release on. Nothing to license across sixty engineers, model spend only. Approved with conditions: managed installation and a stated retention rule.

Treating local inference as one provider among many is the correct design and almost nobody does it. It means my hardware is not a degraded special case bolted on at the end but the same interface everything else uses, which is the difference between supported and tolerated. Permissive terms and a compiled binary complete the picture.

What is absent is a protocol port, so the tool servers I already run cannot attach and every capability has to exist inside this project. Given the licence, that is a fork away rather than a wall.