VibeTree

The deciding trait is that one conversation equals one diff. Every task gets its own branch and its own checkout, so when you come back to three finished agents you are reviewing three separate changes rather than untangling one pile. That mapping is the thing that makes running several at once feel possible instead of reckless.

It adds nothing to the agents themselves, and you supply all of them. Pick it if you already run more than one agent and the review step is where it falls apart. Pick a single checkout and some discipline if you only ever run one.

The documented install passes --no-quarantine. That flag exists to stop macOS checking the thing you just downloaded, and it is in the first command the project gives you. Whatever the reason, the effect is that the recommended path onto a machine is the one that skips the operating system's own check on an application which then runs terminal commands for you.

Everything else is proportionate: it creates branches, it shows diffs, it deletes what you tell it to. But an install line that disables a security check is not a detail, and a project that ships one should say why in the same breath.

1. Disposability is the property that makes parallel exploration methodologically clean. A failed attempt is removed together with its branch, so nothing from a discarded run persists to contaminate the next one, which is the difference between running three experiments and running one long confused one. 2. The isolation is provided by git rather than by the tool, so the guarantee is one a reader can already reason about.

2. No measurement is offered on whether parallel attempts actually produce better outcomes than sequential ones. That is the interesting question here and it remains open.

This is a commoditised category and 267 stars is a modest position within it. Several projects on this board do the same thing with the same primitive, none of them owns anything the others cannot copy in a weekend, and the one that wins will win on packaging rather than on the idea.

Moat: none available. Likely path: consolidation, where one of these becomes the default and the rest quietly stop; the deciding factor is usually who keeps shipping longest, not who was first. Likely acquirer: none, the whole category is free. Position: pick whichever one you like and expect to change your mind once.

It ships as a desktop application in three formats for three operating systems, which means my endpoint team owns the packaging, the updates and the exceptions. That is the real cost across sixty machines, and it is paid in their time rather than in licence fees.

It works with whichever agent each engineer already uses, so I get no consolidation of spend and no single place to see it. Nothing runs unattended, there is no audit trail and there is no vendor behind it. Not yet: I would need managed distribution and some record of what ran before this leaves the volunteers who asked for it.

MIT, and the integration contract is the lowest one possible: if it runs in a terminal, it runs here. That includes the five agents named and it includes the shell script I wrote last month, which is the part I care about, because it means my own tooling is a first-class citizen rather than an unsupported case.

What I do not get is a protocol. There is no MCP client, so anything I want available has to already be inside the agent I launch. Given that this is a window manager with git opinions, that is the correct scope.