Lemon

The deciding trait is where you talk to it. Telegram, Discord, WhatsApp, a terminal or a browser, all reaching the same assistant, which means the thing is available in the places you already have open rather than in a window you have to remember to visit. That changes how often you actually use an agent more than any capability on the list.

What you accept is a very large project maintained by one person. Pick it if the ambient access appeals. Pick something narrower if you want a tool, not a companion.

Per-run supervision is the platform's best idea and its sharpest edge. A crash is survivable because the runtime brings the process back, but the file the previous attempt already wrote, the command it already ran and the message it already sent are not undone by that recovery. Nothing in the row records idempotency, checkpointing, or a rule for resuming mid-task.

What it does right is isolating runs from each other, so one failing task cannot take the rest of the assistant down with it.

1. Building a reproducible arena in which several agents interact is a serious methodological choice, because determinism is what makes a comparison mean anything and almost nobody in this category bothers. 2. The instrument existing without published measurements is the odd part: the hard work is done and the results are absent.

2. That leaves a reader with a testable claim rather than a demonstrated one. The arena is exactly where a sceptic should start, and its presence makes scepticism cheap to satisfy.

The mention count is doing real work here, because thirty-five conversations in a year is more attention than most funded products on this board earn. The gap is that none of it converts: there is no entity, no service, nothing to sell and nobody to sign a contract with, so the attention accumulates and then disperses.

Moat: the platform choice, which keeps casual contributors away and dedicated ones loyal. Likely acquirer: none; a language community absorbs the ideas. Position: use it personally, and do not wait for a company.

Accounting and rate limits arriving as part of the design is unusual and welcome, because it means spend is attributable and a runaway session is bounded before finance notices. That is a control I usually have to build around a tool rather than find inside one.

The rest is absent. No identity integration, no directory sync, no audit export, nothing that runs unattended, and the row records no install steps at all, so putting this on sixty machines is a project of our own. Not yet.

Being a client and a bridge at once is the part nobody else here manages. Servers I already run attach to it, and it presents itself to the other agents on my machine, which makes it a hub instead of another island. Local endpoints are configurable next to the hosted list, so the weights stay mine when I want them to.

Permissive terms on a language built for supervised concurrency, with the source readable and forkable. This is the most bendable thing on this page and I am not being grudging about it.