Twinny

This was a good extension and it is over. The project's own site describes it as archived with the authors moved on, which is about as clear a signal as anyone in this category has ever given, and it removes the need for any judgement on my part about whether it is worth your Tuesday.

Editor extensions decay faster than most software because the editor underneath them keeps changing, so this will break rather than merely age. Pick Tabby if you want local completions and chat that somebody still maintains. Do not adopt this.

The specific failure mode for abandoned extensions is worth stating precisely, because it differs from an abandoned library. The editor publishes new versions on its own cadence, deprecates interfaces, and eventually removes them. A library frozen in time keeps working; an extension frozen in time stops loading, usually after an update the user did not choose to install.

There is no mitigation short of pinning your editor, which creates a worse problem. What it did right: the default configuration pointed at localhost, so it worked offline from the first launch with nothing to sign up for.

1. Completion uses fill-in-the-middle, which requires the served model to support that objective and constrains model choice more than the long provider list implies. 2. Chat context comes from embeddings computed over the workspace locally, so retrieval quality depends on an embedding step running on the same machine as the developer.

2. There is no action stage at all, which places the design firmly before the tool-calling era rather than being a limitation of its execution. No benchmark was published. The local embedding choice was ahead of its moment and is now ordinary.

The interesting event already happened and it was not an acquisition. The people who built this announced a successor and moved, which is how value leaves an open project: not sold, just relocated, with the users left to follow or not. Roughly three thousand six hundred stars accumulated over two years is a modest audience to migrate anyway.

Moat: none existed. Likely path is complete: attention transferred to the next thing, this one left as a marker. Position: none here. Watch what the same authors ship next, because the judgement that produced a localhost default was sound.

This is a two-minute meeting. A tool the publisher has declared archived cannot be standardised across sixty machines, because the first editor update that breaks it lands on all sixty at once with nobody to escalate to and no fix coming. That is an outage we scheduled for ourselves in exchange for nothing.

There is no support channel, no maintenance commitment and no security contact for a component installed on every developer machine. No directory integration, no audit trail, no pipeline role. Onboarding cost is a marketplace click, which is the only cheap thing here. Not yet, and there is no later.

Archived code under MIT is not a tombstone, it is a starting point, and this one is small enough that maintaining a personal fork is a realistic weekend rather than a commitment. The endpoint setting takes anything OpenAI-compatible, so my own server on my own network was always the default rather than a fallback.

What I would actually keep is the local model path and the offline behaviour, and what I would rewrite is everything touching the editor interface, which is the part that rots. I would not recommend it to anyone. I would still have a fork of it on my machine, and I do not apologise for that.