BitFun

The deciding trait is that a task gets an interface instead of a paragraph. Ask for something with shape to it and you get a chart, a board or a form, with the conversation attached to whatever that panel currently shows. If you have ever scrolled back through nine hundred lines of chat looking for a number, you will feel the point immediately.

It is the wrong tool if your work is one file and one question, because building an interface for that is ceremony. Pick it when the task has state worth looking at. Pick a terminal agent when it does not.

Look at the permission surface. This thing drives the browser, the terminal, desktop applications, the filesystem and remote workspaces. Desktop application control is the widest grant on this board and it appears here beside no isolation layer at all, on a machine that also holds your email client. The failure mode is not exotic. It is one confident wrong action with reach.

What it does right is commit. Work lands in a real repository with real history, so the record of what happened survives the session that produced it.

1. The named modes are the interesting part: agentic, plan, debug and deep review are separate documented states rather than one prompt behaving differently, so a user can tell which discipline is being applied. 2. The loop terminates in a test run and a commit, which means verification is part of the design instead of a step left to the operator.

2. No evaluation accompanies this and none is claimed, so nothing invites comparison. The choice to bind a conversation to a live interface state is unusual and, as published, untested.

2,047 stars is meaningful attention for a desktop tool, and there is no company underneath it to convert that attention into anything. The product is free, the model spend goes to a provider, and nothing in the row describes a paid tier, which means the upside today is reputation.

Moat: the interface idea, briefly, until somebody with a design team copies it. Likely acquirer: none directly, though the concept walks into a larger desktop product within a year. Position: adopt personally, and treat the roadmap as a hobby until a business model appears beside it.

One detail here does more for procurement than anything else on the row: multi-device sync runs through a relay we deploy and hold the keys to. That answers the residency question before legal asks it, and it is rare enough that I noticed.

Everything else is missing. No SSO, no SCIM, no audit log, no console for sixty installations, and nothing that executes in a pipeline, so the throughput case stays a story told by volunteers. Approved with conditions: our relay, our keys, and a review of which desktop permissions it is granted.

The extension story is the reason I am here. MCP servers attach, skills load, and the hooks are Codex-compatible, which means scripts I already wrote for one agent fire in this one without a translation layer. Below that I can change the source, and the core carries an MIT badge, though the badge does not cover everything shipped around it.

The key is mine and the provider is my choice at first run. What I cannot do is serve the model myself, so the one part I do not own is the part that thinks.