33,906 stars · Apache-2.0 with a paid ee/ carve-out (Tabby Enterprise License) · v0.32.0 (25 Jan 2026) · Track this in Scout
A coding assistant that runs the model on hardware you control.
▶Repo detailsthe review · specs · pros & cons · install
What it is
Tabby is a self-hosted completion and chat server with an administration screen for a team, plus editor extensions that connect to it. Self-hosted means it runs on a machine of your own choosing rather than as a service somebody else operates. It ships as a Docker container, which is a way of running a program inside its own sealed box so it cannot disturb anything else on the machine.
What it is good for. A team that cannot send source code to an outside provider: regulated work, client code under a confidentiality agreement, or a company rule that forbids it. It also suits anyone who would rather pay once for hardware than carry a per-request bill that grows with use.
- It publishes a real hardware figure, which almost nothing else in this edition does: "approximately 8GB of VRAM for CodeLlama-7B" in int8 mode, with the minimum graphics-card compute capability listed for each precision.
- One container, one command, and an administration screen that handles seats and usage for a team.
- No per-request cost at a provider once the hardware exists, and no code leaves the network.
- OPEN CORE. The licence file is Apache-2.0 with a preamble that carves out the
ee/directory, andee/LICENSEis the Tabby Enterprise License: production use requires an agreed subscription "with a valid Tabby Enterprise license for the correct number of user seats". GitHub shows no licence label for the repository at all as a result, so neither fact is visible from the sidebar. - THE FIRST VISITOR BECOMES THE OWNER. The documented command publishes port 8080 with no credentials set, and the documentation states: "The first registered account after deployment will be the admin account and will be granted the owner role." Anyone who reaches that port before the operator registers takes the instance.
- The README pulls the image
tabbyml/tabbyfrom Docker Hub while the official documentation pullsregistry.tabbyml.com/tabbyml/tabby, which is two different supply chains for the same command. The one published hardware figure is for CodeLlama-7B, a model the quick start never uses. Code last landed 30 June 2026 and the newest release is dated 25 January 2026, so the pace here is slower than in the rest of part three.
- twinnydotdev/twinny
The same self-hostable completion and chat against local models, as a VS Code extension only, with no server component.
Track this in Scout - continuedev/continue
The same idea inside the editor; its README now declares the repository read-only and unmaintained.
Track this in Scout - fauxpilot/fauxpilot
A self-hosted completion server with the same goal; no code since 9 April 2024 and no archived banner.
Track this in Scout
docker run -d --name tabby --gpus all -p 127.0.0.1:8080:8080 \ -v $HOME/.tabby:/data \ registry.tabbyml.com/tabbyml/tabby \ serve --model StarCoder-1B --chat-model Qwen2-1.5B-Instruct --device cuda
