oli

The deciding trait is honesty about its stage. The project tells you plainly that it is very early, which is worth more than a landing page promising otherwise, and it means you can go in curious rather than disappointed. There is a real agent here that searches, edits and runs commands, and you can read the whole thing.

What you should not do is put a deadline behind it. Pick it if you want to understand how these tools work from the inside. Pick Aider if what you actually need is the work finished this week.

The cost of the split lands on the user before it pays off. Two toolchains have to be present, built and version-matched, and a build script plus a run script is the documented way in, so installation is already a step where an early project can lose someone. Any protocol mismatch between the halves shows up as an interface that has stopped responding rather than an error that names itself.

What it does right is state the maturity up front, which is more than most projects at this stage manage.

1. Placing a defined protocol between the reasoning loop and the presentation layer is the right seam, and it is the one most terminal agents fail to draw: it makes the loop exercisable without a terminal, which is the precondition for testing it at all. 2. It also means the interface can be replaced without touching the part that decides anything.

2. The choice is stated as an architecture and accompanied by no evaluation, so the correct reading is that the boundary exists, not that the loop behind it has been measured.

There is no commercial story to read and the project does not pretend otherwise. A single maintainer, an early-stage label in the documentation, no hosted tier and no sponsorship means continuity depends entirely on one person's remaining interest, which is the shortest runway there is and also the cheapest to walk away from.

Moat: none, and none intended. Likely path: it stops when the author's attention moves, as most projects of this shape do, and the ideas resurface elsewhere. Position: read it, learn from it, and do not let a workflow come to depend on it.

A tool whose documented installation begins with a clone and a build script is not something I can roll out. My desktop team ships packages, not compilers, and the platforms it supports leave out a large part of the fleet, so the reachable audience is a handful of engineers who would have installed it themselves anyway.

There is nothing unattended, so it never becomes a step I can measure, and no telemetry or logging I could audit if it did. Onboarding is an afternoon per person. Not yet, and not close.

Most projects that list a local runtime mean it can hold a conversation. This one says tool calling works with a named model, which is the distinction that decides whether an agent is usable offline or merely present offline. That single sentence in the documentation is worth more than a longer provider list somewhere else.

Permissive licence, a compiled core I can read, and a build from source that at least guarantees the binary is mine. There is no protocol client, so anything I have already exposed as a tool server stays out of reach from here.