Neo

The deciding trait is that the work is shown instead of narrated. Plan, active task, tool activity, tests and delegated jobs each have a place on screen, so at any moment you know what it is doing without scrolling back through a conversation to reconstruct it. That single interface decision changes how much of a long run you can actually supervise.

What you give up is the informality of chat, and some people genuinely prefer that. Pick it if you have ever lost track of an agent mid-task. Pick a conversational tool if you never have.

Two claims here need a noun. Containment is asserted without saying what performs it, so a reader cannot tell whether a stray command is stopped by the kernel, by a wrapper, or by a prompt. And the confirmation gate is both optional and partial, which means the set of actions that proceed without asking is decided by the tool rather than by the operator.

What it does right is delegation with evidence. Subagents return findings the coordinator can check rather than summaries it must trust, which is the correct shape for parallel work.

1. Making the investigating subagents read-only and concurrent is the correct decomposition: retrieval parallelises safely, mutation does not, and the design encodes that distinction instead of relying on instruction. 2. Bounding each delegated task means the coordinator's context grows by results rather than by transcripts, which is the difference between a system that degrades slowly and one that degrades suddenly.

2. No evaluation is published and no capability claim requires one, since the argument being made is about structure.

Attention exceeding installs is not automatically bad, it is early. People are discussing it before they depend on it, which for a single-maintainer project is the ordinary sequence and also the point of maximum risk, because interest of that kind decays if the release cadence does. There is no entity and no commercial surface to fund one.

Moat: none, though the interface idea is the sort a larger tool copies rather than buys. Likely path: it stays a well-regarded personal project. Position: adopt for yourself, keep your fork current.

The absence of a plugin ecosystem is worth more to me than it sounds. A fixed tool set is a fixed review: I approve one binary rather than an extensible surface that changes every time a developer installs something. Sessions staying on the machine also removes a data-residency question I would otherwise have to answer.

What is absent is everything else. No identity integration, no provisioning, no central record of what sixty engineers ran, and no unattended mode to measure. Approved with conditions: managed distribution and keys issued centrally rather than exported by developers.

No telemetry, stated plainly rather than buried behind an opt-out I would have to find. That alone buys goodwill, and the permissive licence plus a readable Go tree means when something annoys me I fix it rather than requesting it. One binary, no runtime, no phone home.

Then the ceiling: no MCP client, so my own servers stay outside, and no local endpoint, so inference always leaves the machine whatever else stays on it. For a tool this deliberate about what it does not ship, both feel like decisions rather than oversights, and I still want them.