Copilot-style completion is fantastic right up until your code can't leave the building. Plenty of teams - regulated industries, agencies under NDA, anyone with proprietary client work - simply can't send source to a third-party cloud. Tabby is built for exactly that constraint: a self-hosted AI coding assistant you run on your own hardware, giving your editor Copilot-like completions and chat while every byte of code stays on your network.
A Copilot You Actually Control
Tabby is a completion and chat server plus a set of editor extensions (VS Code, JetBrains, Vim, and more). You run the server; the extensions talk to it. There's no external API in the path, no per-seat SaaS subscription, and no question about where your code went.
Runs on Your Hardware
The server is a single Rust binary (Docker images provided) that serves open code models. Point it at a GPU for speed, or run smaller models on CPU:
# Run the Tabby server with Docker on a CUDA GPU
docker run -it --gpus all -p 8080:8080 -v $HOME/.tabby:/data tabbyml/tabby serve --model StarCoder-1B --device cuda
Repo-Aware Answers
Beyond line completion, Tabby includes an "answer engine" and context features that let it retrieve from your own repositories and docs, so its suggestions and chat answers are grounded in your codebase rather than generic training data.
Open-Core - Which Is Why the License Is "Other"
Tabby's core is open source (Apache-2.0), but the project follows an open-core model: some enterprise features live under a separate commercial license in the repository. That mix is why I've tagged the license here as "Other" rather than a single SPDX identifier - the self-hostable core is genuinely open, but read the licensing if you plan to use the enterprise pieces.
Rough Edges and Limitations
- You run the infrastructure. Self-hosting means you own the GPU, the updates, and the uptime.
- Local models trail the frontier. Open code models are good and improving, but won't match the very best hosted assistants on the hardest completions.
- Hardware sets the ceiling. Completion latency and model size are gated by the GPU you can give it.
- Open-core caveats. Confirm which features are Apache-licensed versus enterprise before you build a workflow on them.
If privacy or compliance rules you out of cloud coding assistants, Tabby is the most established way to give a team that experience on infrastructure you control.
Internal link suggestion: link "editor extensions" to the Continue or Cline post for cloud-optional in-editor AI alternatives.