Tool
Tabby
An open, self-hosted AI coding assistant: code completion, answer/chat-style help in the IDE, and local or on-prem deployment without sending every keystroke to a vendor cloud by default.
What it is
Tabby (TabbyML/tabby) is a self-hosted coding assistant aimed at teams that want Copilot-shaped completions and IDE chat while keeping the server on infrastructure they control. Official messaging covers VS Code, JetBrains-family IDEs, Neovim, and more. You bring (or download) models; Tabby is the serving + client glue, not magic weights.
What it replaces
It replaces GitHub Copilot / other hosted coding assistants for orgs that need on-prem or air-gapped inference and are willing to operate a model server. It does not replace a strong model (local or hosted) — completion quality tracks what you run — and it does not replace code review, tests, or reading diffs before you merge.
License
Most of the Tabby tree outside ee/ is under the
Apache License 2.0. Code under ee/ carries a
separate enterprise license — do not assume Apache-2.0 applies to
everything in the monorepo. Community / self-hosted plans and paid Team/Enterprise
offerings are commercial product choices layered on top of that split. Read the LICENSE
files in the paths you actually ship.
Where it falls over
- Self-hosting means you own GPU/CPU capacity, model updates, and auth.
- Weak or undersized models produce confident wrong code; the IDE still pastes it.
ee/vs Apache-2.0 split surprises teams who only skim the repo root badge.- IDE extension breakage tracks editor APIs — pin versions in regulated environments.
- Enterprise SSO / advanced features may live behind paid plans; verify before promising stakeholders.
Who it's for
Teams that want a self-hosted completion/chat assistant under Apache-2.0 for the core and will budget hardware (or a small GPU box) plus ops time. Wrong default if you wanted zero-ops Copilot and had no interest in model hosting.
Hardware that usually matters
Tabby itself is light; the model server is not. Consumer GPUs are explicitly in scope for many deployments, but VRAM still sets completion latency and model size. Thinner than a full local-LLM hub page — examples only.
- GIGABYTE GeForce RTX 3060 Eagle OC 12GB — entry 12 GB VRAM for local coding models.
- CORSAIR Vengeance LPX DDR4 32GB (2×16GB) — DDR4 system RAM path when the model spills off the GPU on older boxes.
Some links are Amazon Associates. We may earn a commission if you buy.
Related in this cluster
- Continue — BYO-model editor agent; acquisition status worth verifying
- Aider — terminal pair-programmer with git commits
- GitHub Copilot — hosted proprietary baseline many teams compare against
- Cline — open coding agent in the editor
Verdict
Use Tabby when self-hosted completions matter more than a vendor cloud, Apache-2.0
covers the paths you need, and you will size a model honestly. Treat ee/ as
a different license conversation. Stars are not a substitute for reading LICENSE.
← Tools · Home · tabby.tabbyml.com · GitHub