MagenticLite

The trait that decides this one is the pause. It checks in with you before critical steps and every action stays steerable, so filling in a form or working through a research task feels like supervising rather than gambling. Anyone who has watched a browser agent confidently click the wrong button will understand why that is the whole product.

It is an application you sit in front of, not something you embed, and it does not touch a terminal or your repository. Pick it for supervised research and form work on your own machine. Pick Browser Use when you are building the automation into something else.

The risk is institutional rather than technical. It comes from a research group, it has already been renamed once as the successor to an earlier project, and the material describes it in the language of investigation rather than of support. Nothing in that arrangement promises a fix when a website changes its markup and the automation stops working.

Assume no maintenance commitment and pin the version you tested. What it does right: browser sessions are confined to a lightweight virtual machine sandbox, which is the correct default for a component whose entire job is executing instructions found on untrusted pages.

1. Responsibility is divided between an orchestrating model and a separate model trained for browser interaction, rather than a single large model doing both. 2. The stated consequence is that the system runs without frontier-scale compute, which is a claim about a cost-quality frontier. 3. That claim is exactly the kind a benchmark exists to settle, and none is published.

The decomposition itself is principled and should survive model turnover, since either half can be replaced independently. The absence of a measured comparison against a single-model baseline is the missing half of the paper.

There is no company here to underwrite, only a budget line inside a very large one. Research groups publish, get cited, and get reorganised, and the artefacts they leave behind are maintained exactly as long as somebody senior finds them interesting. Nothing is sold, so there is no revenue signal to argue for continuation when priorities move.

Moat: none of its own; the parent's is elsewhere entirely. Likely path is not acquisition but absorption, with the useful pieces surfacing inside a commercial assistant and the standalone application quietly stopping. Position: read the code, harvest the idea, expect the project to end.

Free is not the same as adoptable. There is no support agreement to sign, no named channel to escalate through, and no directory integration, so sixty installations would be sixty unmanaged applications each holding credentials for whatever sites they were pointed at. That is a shadow inventory, not a rollout.

There is no unattended mode either, so nothing it does is centrally logged or reproducible, and an agent operating in a browser with an employee's session is precisely what our controls exist to prevent. Onboarding needs a package manager and an endpoint. Not yet, and probably never at this scale.

MIT from a lab that did not have to make it MIT, installed with uv rather than a shell pipe, and served on a port I choose on the command line. Local endpoints are supported, so both halves of the model split can run against hardware I own, and that is not a footnote here because the design was built for models small enough to do it.

Ten thousand stars means a fork has a constituency if the parent loses interest. No MCP support on either side, which is the one thing I would patch first. Otherwise this is the rare research release I would actually keep.