graff

The deciding trait is size. Three and a half megabytes, no dependencies, one file: it starts instantly, it updates by replacing itself, and there is no runtime underneath it to break on a machine you did not set up. That is a rare shape for a coding agent.

Small also means young, and you will notice the places where a bigger project has already been. Pick it if you value a binary you can carry on a stick and start anywhere. Pick something heavier when the feature you need matters more than the download.

The central claim is unverifiable. The pitch is that you sign in with an AI subscription you already have, and the row records that the documentation does not say which ones. That turns the first five minutes into a coin toss.

The consequence is worse than inconvenience. A provider list that is undocumented is a provider list that can change quietly, and a user has no way to tell whether a sign-in that stopped working is a bug or a decision. What it does right is giving each chat tab its own process, so one wedged conversation does not take the others with it.

1. Most tools that let a model look at a web page hand it pixels and hope. Capturing role, accessible name and selector passes three pieces of structured, machine-checkable information instead, so the reference survives a re-render and can be checked against the live document. 2. The accessibility tree is the correct source for this.

2. The feature is marked experimental and no evaluation accompanies it, which is the honest pairing. The open question is coverage: elements without an accessible name, and applications that rewrite selectors on every build, are the common case on the modern web, and nothing describes the behaviour there.

The language choice is a capital question dressed as a technical one. Zig produces the small artefact everyone admires and shrinks the set of contributors who could maintain it, which for a single-author project matters more than the artefact does. There is no company, no revenue and no funding here.

Moat: none, and the smallest audience on the board makes one unlikely. Likely path: it stays a technically admirable side project, or the author is hired for exactly the taste it demonstrates. Position: enjoy it, and do not put anything on it you would have to rewrite in a hurry.

Distribution decides this before anything else does. The documented install is a script fetched from a release URL and piped into a shell, which is not a channel I can put in front of sixty machines. There is no packaged build listed, so there is nothing to mirror and nothing to pin.

After that the answers are familiar: no single sign-on, no directory sync, no audit export, and it does not run unattended, so it produces no number I can report. Licence cost is zero, which nobody in procurement has ever been persuaded by on its own. Not yet: bring me a signed package.

The source is there and the terms are not. A non-standard licence file means GitHub reports nothing an automated check can read, and I am left guessing whether a fork is permitted at all. That is a strange place for a project this open in every other respect to leave its users.

Everything else is what I want: MCP servers attach, the source reads cleanly, and nothing about the tool asks me to trust a service. Grudging respect, withheld until somebody writes three letters in a LICENSE file and lets me stop guessing.