Tabby

What does Tabby do?

Tabby is "a self-hosted AI coding assistant, offering an open-source and on-premises alternative to GitHub Copilot." It is a server you run, "self-contained, with no need for a DBMS or cloud service", with an OpenAPI interface for integration, and it "supports consumer-grade GPUs." Editor plugins for VS Code, JetBrains and Vim talk to it for completion, chat in a side panel, inline edit through chat commands and generated commit messages.

Completion is retrieval-based: the README describes RAG-based code completion with locally relevant snippets from the language server and recently modified code, and multiple choices in inline completion. Team features live in the enterprise directory: an Answer Engine, team management and access control, a reports tab with team-wise analytics, LDAP authentication, GitLab single sign-on, and GitHub and GitLab integration that indexes merge requests as context. Agent work is pointed at a separate project, Pochi, which connects GitHub issues to tasks and opens pull requests. Models named in the README include StarCoder-1B and Qwen2-1.5B-Instruct in the Docker example, CodeLlama 7B, CodeGemma, CodeQwen and Codestral.

Key facts

• Licence: split. The root LICENSE says everything under ee/ is under ee/LICENSE, the Tabby Enterprise license, and everything outside it is Apache-2.0.
• The enterprise licence permits production use only with a valid paid seat licence under Tabby's subscription terms, allows copying and modifying for development and testing without a subscription, and forbids distributing or selling the software.
• What sits under ee/: tabby-webserver, tabby-ui, tabby-db, tabby-schema, tabby-email and related crates, which is the web server and team layer.
• Self-hosted server via Docker (tabbyml/tabby) with a GPU in the example, or built from source with Rust; deployable on any cloud through SkyServe.
• Editors: VS Code, JetBrains Marketplace, and Vim plugins.
• RAG-based completion with local snippets and multiple inline choices; chat and inline edit; commit messages.
• Team features: Answer Engine, team management, analytics reports, LDAP, GitLab SSO, GitHub and GitLab integration.
• Latest tagged release when read: v0.32.0 on 2026-01-25; last push 2026-06-30; 33,884 stars on 2026-09-18.

What does it replace, and where does it fall short?

Tabby replaces GitHub Copilot for a team that must keep code and completions inside its own network: one server, its own GPU, its own models, and plugins in the editors people already use. It covers the completion half of Cursor and Windsurf the same way. It is the self-hosted pick on the open-source alternatives to GitHub Copilot page.

Where it falls short: the web server and every team feature are under the enterprise licence, which requires a paid seat licence for production use, so the open-source part is the core rather than the product a team sees. Agent features are delegated to a separate project. Release activity is slow, with the last tag in January 2026 and the last push in June. For an open-licensed completion tool in VS Code, twinny is the simpler pick.

How does Tabby run?

Run the Docker image tabbyml/tabby with a GPU attached, a completion model such as StarCoder-1B and a chat model such as Qwen2-1.5B-Instruct, on port 8080; the README's example passes --device cuda. Apple M1 and M2 Metal inference is supported. Build from source with cargo, protobuf and, on Ubuntu, libopenblas-dev. Then install the VS Code, JetBrains or Vim plugin and point it at the server. TabbyML sells the enterprise licence for the ee/ features; the README does not describe a hosted tier.

Who is Tabby for?

An engineering team that wants Copilot-style completion with nothing leaving its network and is either fine with the Apache-2.0 core or willing to buy enterprise seats for the team layer. An individual developer should look at twinny or Kilo Code first.

What limits does the README state?

None listed in the README. The enterprise licence's production-use restriction, quoted above, is the limit to know.

Questions people ask

Is Tabby open source?

Partly. The root LICENSE applies Apache-2.0 to everything outside the ee/ directory, and the Tabby Enterprise license to everything inside it, including the web server, the UI and the database layer. Because the product a team deploys includes ee/, this site labels Tabby source available; the core alone would be open source.

Can Tabby be used in production for free?

The Apache-2.0 core, yes. The ee/ components may only be used in production with a valid paid Tabby Enterprise licence under Tabby's subscription terms, and may be copied and modified for development and testing without one. The licence also forbids distributing or selling that software.

Does Tabby need a GPU?

The README's Docker example attaches a GPU and passes --device cuda, and it says Tabby supports consumer-grade GPUs. Apple M1 and M2 Metal inference is also supported. CPU-only serving is not described in the README.

How does Tabby compare with GitHub Copilot?

Tabby's README states its own position: an open-source, on-premises alternative to GitHub Copilot. You run the server and the models; the plugins deliver completion and chat in VS Code, JetBrains and Vim. Copilot is hosted by GitHub with an agent and repository-wide features. The open-source alternatives to GitHub Copilot page compares the options.

Which models does Tabby run?

Any model it supports from its own registry; the README names StarCoder-1B and Qwen2-1.5B-Instruct in the quick start, and CodeLlama 7B, CodeGemma, CodeQwen and Codestral in its updates, with Llamafile deployment as an option. Each model carries its own licence.

Sources

• Tabby README and root LICENSE: github.com/TabbyML/tabby, read 2026-09-17.
• Tabby Enterprise license: ee/LICENSE, read 2026-09-17.