Autohand Code CLI

The deciding trait is that everything happens at a single prompt. Mentions pull a file into the conversation, slash commands do the rest, and session history means yesterday is still there this morning. It is a familiar shape done tidily, and the terminal rendering is pleasant in the way that matters when you stare at it for six hours.

What you do not get is a community. Nobody has written the blog post about the thing you are stuck on. Pick it if a clean interactive loop is enough. Pick Claude Code if you want company.

Two claims sit badly together. Risky operations require confirmation, and the intended uses include an automated pipeline. Nothing in the row says which side gives way: whether the gate blocks forever, whether it is disabled by a flag, or whether unattended mode simply approves everything. There is no container boundary underneath either, and it can commit and branch on its own.

What it does right is naming permissions as a first-class surface rather than burying the question in a configuration file nobody reads.

1. Manual mention-based context is an underrated design. The model sees what a person named, so a wrong answer is traceable to a wrong selection instead of to an opaque ranking function nobody can inspect. 2. The cost is recall: a file the user did not think of is invisible, and no fallback index is described that would catch the omission.

2. No benchmark is published and no capability figure is asserted, which is the honest pairing. A tool making an interface argument does not owe anyone a score.

A named vendor with a marketing site behind a free tool usually means the tool is the funnel and the revenue arrives later. Here the later part is missing: no hosted service, no team edition, nothing gated. That is either admirable or unfunded, and the star count is too small to tell which.

Moat: none visible; the interface ideas are copyable in a weekend. Likely acquirer: nobody yet, and the realistic exit is the founders taking the brand somewhere else. Position: use it freely, and do not wait for the enterprise plan.

Headless operation is the reason this reaches my desk at all. A tool that runs in a pipeline can be given a scope, a budget and a report, which is more than most desktop agents ever earn. The licence costs nothing across the whole team, so the only line item is model spend.

Everything governance-shaped is absent. No single sign-on, no directory sync, no central record of what ran and against which repository. The row also lists no installation command, so packaging it for sixty machines is work we would do ourselves. Not yet.

The licence is the good news and it is real news: permissive terms mean the fork stays available whatever the company decides later. Switching providers without restarting is the kind of small freedom I notice, since it lets me route a cheap model at the boring half of a session.

The rest is closed by omission. My own hardware is not a supported destination, and the tools I have already built cannot be attached over a protocol, so the extension story ends at whatever the skills mechanism accepts. Readable source, narrow doors.