Tabby is a self-hosted, open-source AI coding assistant that acts as an on-premises alternative to GitHub Copilot, aimed at developers and engineering teams that want code completion and chat without routing source code through a third-party cloud service.

What it is

Tabby is an AI coding assistant written in Rust that runs inside the user's own infrastructure. It combines inline code completion with a chat side panel and an Answer Engine, and it integrates with editors through shipped plugins, including the Visual Studio Code extension published as TabbyML.vscode-tabby. The project is self-contained: the README states that it needs no DBMS and no cloud service, so the assistant and its model serving can sit on hardware the operator controls. It is distributed through the tabbyml/tabby Docker image, and the project maintains documentation, a Slack channel, and a public roadmap at tabbyml.com.

The concrete problem it solves is source-code egress. Commercial assistants typically require sending code context to a vendor's hosted model. Tabby replaces that arrangement with a local deployment: the same completion and chat workflows, but the model inference and the indexed context stay on hardware the operator owns. It supports consumer-grade GPUs, which lowers the entry cost for teams that want private inference but do not have datacentre-class accelerators. Teams that already run a cloud IDE or an internal developer platform can reach Tabby through its OpenAPI interface rather than through a vendor-specific integration.

Key capabilities

• Code completion and chat delivered through editor plugins, including TabbyML.vscode-tabby, with a chat side panel and an @ menu for mentioning files as chat context (v0.27).
• Answer Engine, described as a central knowledge engine for internal engineering teams that integrates internal data; v0.28 turns Answer Engine messages into persistent, shareable Pages, and v0.20 adds switching between different backend chat models.
• OpenAPI interface for integrating Tabby with existing infrastructure, such as a cloud IDE.
• Self-contained operation with no DBMS or cloud service dependency, plus support for consumer-grade GPUs.
• Context enrichment from real repositories: v0.30 indexes GitLab merge requests as context, v0.29 adds custom documentation through REST APIs, and Codestral integration was announced in 2024.
• LDAP authentication and improved notifications for background jobs (v0.24).
• Llamafile deployment integration (v0.21) and Pochi agent integration that connects GitHub issues to Pochi tasks and opens pull requests with a breakdown of CI, lint, and test results ([email protected]).

Who uses it and how

• Engineering teams under constraints that forbid sending source code or prompts to an external cloud service, who deploy Tabby on their own hardware instead.
• Small teams and individual developers running inference on consumer-grade GPUs, since the README explicitly targets that hardware class rather than datacentre accelerators.
• Platform teams wiring the assistant into an existing cloud IDE or developer portal through the OpenAPI interface.
• Internal engineering teams that need answers grounded in private documentation, using Answer Engine with custom docs loaded via REST APIs (v0.29) and shared threads surfaced on the main page (v0.19).
• Teams on GitLab that want merge requests indexed as completion context rather than assembling that context by hand.

Getting started

Run the tabbyml/tabby Docker image, which is the distribution path the README badges point to, and pair it with the editor plugin such as TabbyML.vscode-tabby. Documentation and deployment guidance live at tabbyml.com, with Llamafile listed as an alternative deployment route.

How it compares

The README positions Tabby directly as an open-source, on-premises alternative to GitHub Copilot. Where Copilot is a hosted commercial service, Tabby is self-hosted and its code is published openly, so the operator controls the model, the indexed context, and the data path rather than a vendor. That changes the cost model from a commercial subscription into infrastructure the operator supplies, chiefly GPU capacity, and it removes the requirement to send code to a third party.

When to use it — and when not to

A self-hoster must run the tabbyml/tabby container and provide a GPU, though a consumer-grade one is enough, and anyone using Answer Engine over internal documentation must load that content through REST APIs, with LDAP configuration needed if directory authentication is wanted. The database requirement is notably absent, which keeps the operational surface smaller than many self-hosted assistants. Teams that want a fully managed service with no hardware and no operations should not pick Tabby, and the registry reports the licence as NOASSERTION, so the licence terms should be confirmed before adoption.

docker run -it \
  --gpus all -p 8080:8080 -v $HOME/.tabby:/data \
  tabbyml/tabby \
  serve --model StarCoder-1B --device cuda --chat-model Qwen2-1.5B-Instruct

git clone --recurse-submodules https://github.com/TabbyML/tabby
cd tabby

# For MacOS
brew install protobuf

# For Ubuntu / Debian
apt install protobuf-compiler libopenblas-dev