Raccoon

What decides it in daily use is scope. This is completion, chat and an in-file edit command, and nothing about it pretends otherwise, so what you get is the 2023 shape of an assistant at a time when the rest of this board edits across files and runs your tests. That is not a criticism of the execution, it is a description of the ambition.

You will outgrow it in a month. Pick it if the choice is not yours. Pick Supermaven if it is.

The gap is disclosure. Completion as you type means source leaves the editor continuously, and the only published material is a marketplace entry that identifies neither the model nor what happens to the text after it arrives. A reader cannot determine whether their code trains anything, because nothing addresses the question.

What it does right is not overreaching. There is no autonomous mode and no shell access, so the failure modes are limited to bad suggestions rather than bad actions, and a bad suggestion is something a reviewer catches.

1. Documentation is thin to the point of absence. There is no technical page, no repository and no description of how the edit command decides what to rewrite or how the result is applied. 2. That makes every capability here a claim rather than a demonstration.

2. No benchmark accompanies the product, which at least spares the reader a methodology dispute. The honest summary is that there is nothing to audit, and for a tool that reads source continuously, absence of documentation is itself a finding.

Free with no upgrade path means this exists to keep the vendor visible in a category it does not lead, rather than to earn anything. Products in that position have no internal champion when budgets tighten, and the marketing page carries none of the signals, a roadmap or a changelog, that would suggest active investment.

Moat: none; the parent's model business is the asset and this is a poster for it. Likely path: quiet deprecation, or a relaunch under a different brand. Position: pass, and revisit only if the vendor gives it a price, which would at least prove someone owns it.

The price makes the finance conversation trivial and makes every other conversation harder, because there is nobody to negotiate with. Installation is per developer from a marketplace, so sixty machines is sixty independent decisions I cannot enforce or revoke centrally.

Nothing here runs in a pipeline, so it produces no records and improves no throughput I can measure. Onboarding is five minutes, which is the one honest advantage. Not yet, and in this form not ever: I cannot approve continuous source egress to a service with no published terms.

This is the closed box in its purest form. The licence is proprietary, the model is unnamed, no key of mine can be substituted, and the row records inference happening somewhere that is not my machine. There is no extension surface at all, so the tool does exactly what it does forever.

The configuration I get is whichever toggles the extension settings expose, which is not configuration, it is decoration. No fork, no offline mode, nothing to script. I have no grudging respect available for this one.