Sinew

The deciding trait is the three modes. Act, Goal and Plan are separate states you enter deliberately, so asking for a plan is a mode rather than a phrasing you hope the model honours. Anyone who has watched an agent start editing during what was meant to be a discussion will recognise why that separation is worth a menu.

The cost is that this is a desktop application competing with tools that live where you already work, and it is very young. Pick it if explicit modes appeal. Pick a terminal agent if another window does not.

The dealbreaker is a credential path, not a code path. Signing in through authentication flows that a vendor built for its own clients puts the user's subscription in a category the vendor did not agree to, and the consequence lands on the account rather than on the tool. The row states this plainly, which is creditable and does not make it safer.

What it does right is offer the alternatives beside it. A plain key and an aggregator route are both documented, so the risky path is a choice rather than the only door.

1. Making the tool surface a configured subset rather than a fixed set treats capability as a variable, which is the correct treatment: a smaller surface reduces both selection error and prompt length, and the operator decides the trade rather than the vendor. 2. Sourcing instructions from two committed files puts the system prompt under version control, so a behavioural change has a commit behind it.

2. No evaluation is published and none is claimed. The argument is about control, which is inspectable without one.

Desktop is where independent tools go to be undiscovered. There is no package registry pulling users in by accident, no marketplace listing, and every install is a deliberate download of a binary from a stranger. Sixty-four stars after that funnel is roughly what the funnel predicts.

Moat: none, and the differentiators are configuration choices any competitor could adopt in a release. Likely path: it stays a personal project, and the ideas surface in something with a distribution channel. Position: run it if you like it, expect nothing to be built on it.

Configurability is a virtue for one engineer and a support problem for sixty. If every installation can be reshaped locally, then a bug report describes a configuration I cannot reproduce, and a behaviour I approved on one machine is not the behaviour running on the other fifty-nine. There is no central policy surface to prevent that.

Add the usual absences: no identity integration, no provisioning, no audit record, no unattended run to measure. Sixty desktop installs to keep patched. Not yet, and I would need a locked baseline configuration before revisiting.

This is built the way I would build it. Permissive licence, a Rust core I can read, MCP so my existing servers arrive as tools, and a provider layer that treats the model as swappable rather than as a partnership. Nothing about the model choice is anybody's decision but mine.

Two omissions bother me. No documented install path, so I compile it myself, which is fine for me and disqualifying for most. And no local endpoint, so the one component I cannot bring in-house is the inference, in a tool otherwise entirely under my control.