vix

The deciding trait is that it will argue with you before it writes anything. It puts the plan on a canvas and walks you through it out loud, and you push back on the shape of the work while it is still a drawing rather than three hundred lines of diff. Most agents give you a bulleted plan you skim and approve out of politeness.

Whether you want that depends entirely on how you think. Pick it if you review plans better than you review diffs. Pick something quieter if a voice walkthrough sounds like a meeting you did not schedule.

The agent reads and edits minified code through a virtual filesystem. What it sees is not what is on disk, and every edit has to be mapped back through a Tree-sitter view to the real file. That mapping is the failure surface: a stale parse, an unusual construct, a file the grammar handles badly, and the change lands somewhere adjacent to where it was meant to.

Nothing in the row describes what happens when the round trip fails, and there is no git integration to notice. The sandbox is real and the mechanism is not named, which is the second thing to ask about.

1. The token claim is 20 to 50 percent, and the comparison behind it is described by its own authors as an observation rather than a benchmark. Crediting them for saying so is the right response; treating the range as a measurement is not. 2. A saving of that size depends entirely on the corpus, since the technique is a compressed representation of source and its benefit scales with how verbose the source was.

2. What would settle it is a published harness and a fixed task set. Neither exists, so the number is a report from one machine.

The feature list is the tell. Five distinct product ideas live in this one repository, from a project with 274 stars, which is more surface than a small team can defend, and surface area is the thing that quietly consumes a maintainer's year.

Moat: the efficiency angle, if it holds, because cost per task is the one axis buyers will eventually shop on. Likely path: the breadth narrows to whatever people actually use, or it stalls under its own weight. Likely acquirer: none at this size. Position: interesting thesis, immature vehicle, check back after one more release.

The part that concerns me is that it writes its own scheduled jobs, watchers and alerts. An agent that can create recurring work on a developer's machine is a source of activity nobody in my organisation approved, and there is no console anywhere that tells me what is scheduled across sixty laptops.

Cost is zero and macOS and Linux cover most of my estate. Everything else procurement asks for is absent: no identity, no audit, no retention statement, no support. Not yet: I would want the scheduler disabled by policy and a central view of what it is running before this goes past a single volunteer.

AGPL-3.0, MCP servers attach, and the part I would actually use is the pipeline format: multi-phase workflows defined as JSON with agent, bash and tool steps, plus templating, branching and history forking. That is a program I can version and diff, not a wizard I have to click through, and bash steps mean anything on my machine is reachable from inside a phase.

The provider list is four hosted names and no local endpoint, which is the wall. Everything above the model is mine to shape; the model itself is somebody's API, and for a tool this configurable that is a strange place to stop.