Sculptor

You will get the most out of this when a task has more than one reasonable approach. The trait that decides it in daily use is that each agent gets its own branch and its own workspace, so running three attempts is normal rather than reckless, and the thing you review at the end is a pull request on GitHub rather than a pile of edits in your checkout.

Pick it if comparing attempts is how you like to work. Pick Claude Squad for the same idea in a terminal, or a single-agent tool if you would rather spend your attention on one attempt done well.

Two documents from the same vendor disagree about the safety model, and the more optimistic one is the page a buyer reads first. A worktree separates files and shares everything else, including the shell, the network and your credentials, so a user who believed the marketing has a materially different threat model from the one they actually have. The vendor labels the whole thing an experimental research preview, which is honest and does not resolve the contradiction.

What it does right: changes are reviewed live as diffs before they go anywhere, so the human checkpoint is built into the loop rather than offered as an option.

The integration deserves attention. 1. The underlying agent is run as a structured process over a documented control channel rather than by parsing its display, so the wrapper reads events instead of pixels. 2. Two tools in the agent's namespace are replaced with the host's own implementations, which lets the surrounding application own the moments where a human is asked a question or a plan is formed.

That second choice is the substantive one: it means the supervision layer is inside the agent's tool loop rather than bolted around it. No evaluation is published, and the vendor's own preview labelling is the correct signal about maturity.

Read the distribution before the features. Roughly two hundred stars in a category where competitors count them in thousands says this has not found an audience, and there is no pricing page at all, which means nobody has been asked to pay and therefore nobody has validated that they would. A research organisation publishing tooling is a legitimate activity and not a business.

The exit here is a paper or a pivot, not an acquisition; the interesting asset is the supervision technique rather than the application. Position: adopt as a source of ideas, and do not put a team's workflow behind a lab's side output.

A tool that opens pull requests and then sends an agent to fix failing checks is a tool acting on my repositories, and the row gives me no single sign-on, no SCIM and no audit trail to attach those actions to a person. That is the whole conversation. Distribution narrows it further: Apple Silicon and Linux only, so a third of my engineers cannot install it, and each user needs their own subscription with the upstream vendor.

Onboarding is short, which is not the problem. Not yet. Bring me identity and a Windows build, in that order.

MIT is a real gift on a desktop application, because it means the supervision code is readable and a fork is legitimate, and the repository is public rather than a marketing mirror. That is where my enthusiasm stops. There is no MCP support in either direction, so the servers I run are simply not part of this, and the agent underneath is one vendor's client authenticated with my own login.

Nothing runs on my hardware. I would clone this to study how it takes over an agent's tool loop, and I would not make it my daily driver.