Lemon AI

The deciding trait is the element editor. When it has produced a page and one section is wrong, you click that section and it rewrites that section, which is a smaller and far more useful interaction than writing a paragraph explaining which paragraph you meant.

Everything else about it asks for patience: this is a local system doing work that hosted products do on much larger machines, and you will feel the difference on anything long. Pick it if you want the whole loop on hardware you own. Pick a hosted platform if speed was what you were buying.

The dealbreaker is in the install line. Mounting that socket gives a process inside the container the ability to start further containers with whatever settings it likes, including ones that mount the host filesystem. The boundary being advertised is a boundary the setup instructions remove.

It is a known pattern and a known escape, and the documentation presents it as the normal way to run the product. Nothing describes an alternative deployment. What it does right is keeping code execution off the host filesystem by default, which is the correct instinct undone by the plumbing beneath it.

1. Putting reflection inside the isolated environment means the critique reads the same filesystem the work produced rather than a summary of it. That is the correct place for it: a reflection step that cannot observe the artefact is reviewing a description. 2. Naming the phases makes the loop auditable.

2. Nothing measures any of it. There is no evaluation of whether the reflection step improves outcomes, which is the one claim in this architecture a benchmark could settle cheaply and nobody has. The design reads as deliberate and the evidence for it is a paragraph of prose.

Naming a funded competitor is a distribution strategy and a good one, because it borrows an existing category instead of explaining a new one. What it cannot borrow is the revenue. The competitor sells a subscription, this ships a container, and the gap between those two is the entire business question.

Moat: local-first positioning, which is real and narrow, because the buyers who insist on it are the ones least likely to pay for software. Likely path is a paid edition or a slow fade. Position: run it if the requirement is genuine, and do not plan around the roadmap.

The prerequisite is my whole objection. A large part of my engineering organisation works on Windows, and for them this means installing a subsystem and a container runtime on a managed laptop, which is a change request, a security exception and a support queue rather than an install.

Beyond that there is nothing to govern with: no accounts, no roles, no retention policy, and no record of what any instance did. It does not run in delivery, so it produces nothing I can measure. Not yet, and I would need a server deployment my team operates rather than sixty local ones.

This is my recurring complaint. A licence that takes a well-understood permissive one and adds terms is a licence nobody has read, so every question about what I may do with a fork ends at somebody's email address rather than at a document. The name it borrows does not carry over.

The rest is close to what I want. Ollama and vLLM are documented targets and the model list is full of open weights, so nothing has to leave the machine, and the whole thing arrives as a container rather than as an argument with a package manager. Read the terms first.