dalao.ai

Tabby

Self-hosted AI coding assistantOpen source

About Tabby

No commits to the main repository since 2026-07-01 Commit history

Tabby is a fully self-hosted AI coding assistant focused on code completion: a Rust server, web console and VS Code, JetBrains and Vim/Neovim plugins give teams GitHub Copilot-style completion with code and models kept in-house. TabbyML, Inc., co-founded by ex-Google engineers Meng Zhang and Lucy Gao, announced a $3.2M seed in October 2023. Most code is Apache-2.0; ee/ is licensed separately. Chat and knowledge left main in January 2026. As of 2026-10-26 main's last commit was 2026-07-01 and the latest release was v0.32.0 (2026-01-26). The team launched a new coding agent, Pochi, in July 2025.

EditorialUpdated 10 min read

The Tabby interface

What it does

Plenty of companies want AI completion for their engineers but can't send private code to a third-party cloud, especially in regulated fields like finance and healthcare. Tabby runs the whole completion service in-house. The server handles inference and indexing itself, needs no external database or cloud service, exposes an OpenAPI interface for plugging into existing infrastructure such as cloud IDEs, and runs on consumer GPUs. Those are the three key features listed at the top of the README.

The approach is end-to-end optimization rather than a thin wrapper around a model. The docs' principles page says the IDE plugins use streaming, cancellation and adaptive caching to return completions in under a second, and the server parses relevant code into Tree-sitter tags to build prompts. Since v0.3.0 in October 2023, repository-level retrieval has been on by default, pulling related functions and examples from other files in the same repo as context. Later releases added indexing for GitHub and GitLab code and merge requests, and ingesting internal docs through a REST API.

Who makes it: the company is TabbyML, Inc. (the copyright holder in the LICENSE), and its cofounder is Meng Zhang. In an interview published by WVV Capital, Meng describes spending more than eight years at Google, the last four on generative AI models. Lucy Gao spent four years on computer vision on the same Google team, then joined TikTok as one of its early product team members. The two began serious talks about the idea in late 2022, and the project launched publicly in April 2023. In October 2023 TechCrunch reported a $3.2M seed round from Yunqi Partners and ZooCap. WVV Capital announced an investment in November 2024.

The product line has split since 2025. In July 2025 the site launched Pochi, a coding agent billed as a full-stack AI teammate, and the README's latest news item (2025-12-12) is about Pochi. Meanwhile PR #4425 ("sunset knowledge and chat features"), opened on January 20, 2026 and merged into main on 2026-01-24 00:24 (UTC+8), turned off Tabby's chat feature and removed the knowledge pages and search pages. After that merge, what remains on main is code completion and code-context indexing.

The release history, per GitHub Releases: the first stable release, v0.0.1, came on 2023-08-31. v0.1.1 that September added Metal inference for Apple M1/M2, and v0.7.0 in December added team management and secured access. In 2024, v0.9.0 in March brought a full admin UI, v0.10.0 in April added team usage reports, and v0.13.0 in June introduced Answer Engine, a Q&A engine for internal engineering teams. In 2025, v0.24.0 in February added LDAP sign-in and v0.30.0 in July could index GitLab merge requests. When the seed round landed in October 2023, TechCrunch put the GitHub star count at about 11,000. After v0.30, the README news list's next entry is a Pochi item dated 2025-12-12.

Tabby is a self-hosted AI coding assistant, offering an open-source and on-premises alternative to GitHub Copilot.

Tabby README

Key features

An all-in-one server. Written in Rust, with llama.cpp for inference and Tantivy for the search index. It ships a web admin console with user and team management, usage reports, a code browser and an activity log. Data is stored in ~/.tabby by default, and there's no separate database to install. The repo also includes Ollama bindings (ollama-api-bindings), so teams already running models in Ollama can reuse them.

Repository-level context. Tabby connects to repos from GitHub, GitLab and self-hosted providers, indexes multiple branches (new in v0.32.0), and uses GitLab merge requests and internal docs pushed through the REST API as context for completion and Q&A. Completion also draws on declarations from the local LSP and recently edited code.

Many models, many kinds of hardware. It supports CUDA, ROCm, Vulkan, Apple Metal and CPU. int8 needs an NVIDIA GPU with compute capability 7.0 or higher, or exactly 6.1; float16 needs 7.0 or higher and bfloat16 8.0 or higher. Inference defaults to int8, and per the FAQ CodeLlama-7B needs about 8 GB of VRAM. The same model on an AMD Radeon RX 7900 XTX under ROCm also shows about 8 GB in the monitoring tools, though the FAQ says ROCm limits are largely untested. Each instance uses a single GPU, so multiple GPUs mean multiple instances, each assigned a card via CUDA_VISIBLE_DEVICES or HIP_VISIBLE_DEVICES. Completion and chat models are set separately, and either can point at a remote OpenAI-compatible API.

IDE coverage. Official plugins cover Visual Studio Code (and VSCodium), the IntelliJ platform and Vim/Neovim. The VS Code extension offers multiple completion choices, auto-generated commit messages, inline edits and @-mentioning files in the side panel.

Paid-tier features. Per the pricing page comparison table, features reserved for paid tiers include SSO, LDAP (v0.24), custom branding (v0.31.0, enterprise only), enforcing an IDE telemetry policy and restricting sign-up to an authentication domain.

How it compares

GitHub Copilot: Individual Pro is $10 a month, and GitHub's pricing page says code completions are unlimited on every paid plan, while the Free plan is limited to 2,000 a month (as of 2026-10-25). Copilot is a cloud service; Tabby's distinction is that code stays inside the network, which is decisive only where compliance forbids sending code out.

Continue plus Ollama: Continue is an open-source IDE extension and Ollama runs models locally. Together they also provide local completion and chat, which suits individuals and teams happy to assemble the pieces. Tabby can itself use Ollama as a model backend. Its edge is a central server: accounts, teams, usage reports and repo indexing are managed in one place.

Coding agents like Cline and Aider: they're good at changing code across files and running commands on request, with local or cloud models. Tabby covers the inline-completion layer. The two run side by side; neither replaces the other.

Commercial private deployments like Tabnine and Sourcegraph: better for large enterprises that want a vendor contract and will pay for on-prem deployment. Sourcegraph's pricing page lists Enterprise from $16K a year as a minimum annual contract, with credits for AI features included (as of 2026-10-25); Sourcegraph's blog says Cody Free and Pro were discontinued on 2025-07-23. Tabnine was acquired by the testing-platform company Tricentis on 2026-07-30, and its old pricing page now redirects to a Tricentis contact-sales page. Tabby's advantage is that most of its code is open source and can be audited and modified, and the Community edition is free for up to 5 users.

Who uses it, and who it's for

Adoption figures as of 2026-10-25: the main GitHub repo has 33,921 stars and 1,791 forks. The VS Code extension shows 85,640 installs and an average of 4.5 from 13 ratings (VS Code Marketplace), and the JetBrains plugin shows 35,515 downloads (JetBrains Marketplace). The GitHub contributors API ranks Meng Zhang first, with 1,903 commits.

Who it's for

Teams whose code can't leave the network but who still want AI completion for their engineers. Server, models and index all run on the team's own hardware, and the plugins talk only to the team's Tabby address.

Small teams with a spare GPU. The docs recommend an NVIDIA T4, 10-series or 20-series card, or Apple silicon such as an M1, for 1B to 3B completion models. The Community edition is free for up to 5 users.

People who want control over which model runs and want to see its license. The official model registry lists completion models such as StarCoder, CodeLlama, DeepSeekCoder, CodeGemma and Qwen2.5-Coder, each with its license, and local models that follow the spec can be loaded.

Companies on GitLab or self-hosted Git that need LDAP or SSO. Those integrations arrived from v0.12 and v0.24 on, and v0.32.0 added generic OAuth.

Not for

Anyone after an agent that edits many files and runs commands, like Cursor or Claude Code. Tabby focuses on completion, chat has been removed from main, and the agent work has moved to the team's other product, Pochi.

Solo developers who don't want to manage GPUs and model updates. GitHub Copilot is a hosted subscription with no server or model to run.

Teams that want several GPUs to serve one large model. The FAQ says each Tabby instance uses a single GPU; multiple cards mean running separate instances.

Large buyers who need a dependable long-term roadmap. As of 2026-10-26 the main repo's latest commit was on 2026-07-01, and the latest stable release is still v0.32.0 from January 2026.

Getting started

The Docker command, with the NVIDIA Container Toolkit set up, is: docker run -d --name tabby --gpus all -p 8080:8080 -v $HOME/.tabby:/data registry.tabbyml.com/tabbyml/tabby serve --model StarCoder-1B --chat-model Qwen2-1.5B-Instruct --device cuda. On systems with SELinux, the volume mount needs :Z. On Apple silicon, Homebrew installs it, and the docs include guides for Linux, Windows and WSL2. For GPUs, the model registry recommends at least a T4, 10-series or 20-series card, or Apple silicon such as the M1, for 1B to 3B completion models, and a V100, A100, 30-series or 40-series card for 7B to 13B. Chat is less latency-sensitive, and the model registry recommends at least 1B parameters for chat models. Chat models in the registry include Mistral-7B, the Qwen2.5-Coder series (Apache 2.0), CodeGemma-7B-Instruct (Gemma License) and CodeQwen-7B-Chat (Tongyi Qianwen License), and their licenses differ for commercial use.

Once it's running, the console is at http://localhost:8080. The first account registered becomes the admin. After signing in, the home page shows the server address and token the plugins need, and teammates can be invited from that page. The plugin installs with ext install TabbyML.vscode-tabby in VS Code, or from a Tabby search in the JetBrains marketplace, and then takes the address and token.

Two notes. First, chat has been removed from main, and --chat-model can be left out for a completion-only deployment. Second, usage statistics are sent to TabbyML unless TABBY_DISABLE_USAGE_COLLECTION=1 is set before starting.

Pricing & license

Prices on the pricing page, as of 2026-10-25: Community is free for up to 5 users, deployed locally, with code completion, Answer Engine, inline chat and context providers. Team is $19 per user per month for up to 50 users. Enterprise is custom-priced, with unlimited users, SSO and a dedicated Slack channel. The page also has a Tabby Cloud usage-based pricing section that actually describes Pochi: billing by the token cost of the models used, with $20 in free monthly credits. The Pochi section also says Tab completion is always free, with no usage limits.

The license has two layers. The root LICENSE says everything under ee, including tabby-webserver, tabby-ui and tabby-db, falls under the Tabby Enterprise License. Production use requires following the subscription terms and holding a valid enterprise license for the right number of seats; copying and modifying for development and testing is free. Third-party components keep their own licenses, and everything else is Apache-2.0. GitHub reports the repo's license as NOASSERTION, because the root LICENSE combines two licenses in one file. The admin console and web UI themselves sit in the ee directory.

Risks & limits

First, the latest commit and release. The last stable release is v0.32.0 (2026-01-26), followed by an internal next-alpha prerelease. Per the GitHub API, as of 2026-10-26 the latest commit on main is 21b2904, dated 2026-07-01, which reverted an outside contribution (#4448, adding Avian as a model provider). The newest stable entry on GitHub Releases is still v0.32.0, published 2026-01-26 01:07 (UTC+8). The Pochi repo (Apache-2.0, created 2025-08-20, docs at docs.getpochi.com), by contrast, was still being pushed to on 2026-10-23. As of 2026-10-25 the GitHub contributors API shows 1,903 commits from Meng Zhang in Tabby and 2,614 in Pochi.

Second, chat and knowledge features have been removed. Both are gone from main, yet as of 2026-10-25 the site and pricing page still describe Answer Engine and inline chat.

Third, usage statistics are on by default. The docs say non-sensitive usage data is collected by default. The listed fields include the completion and chat model names, device type, CPU architecture and core count, CUDA devices, version, and whether the web server is enabled. It can be turned off with an environment variable.

Fourth, hardware and model licenses. Each instance uses a single GPU, and some models in the registry have restrictive licenses; Codestral-22B, for example, is under Mistral's non-production license.

Our verdict

For teams whose code can't leave the network and that have 5 people or fewer, or that pay $19 per user per month for up to 50, Tabby is an open-source, self-hosted completion option that deploys with a single Docker command. But chat has been removed from main, as of 2026-10-26 the main repo's latest commit was on 2026-07-01, and the latest stable release is still v0.32.0 from 2026-01-26. For new projects it is now positioned as a completion service only; alternatives include Continue with local models and commercial on-prem products.

Why it stands out

Tabby is a coding assistant you can run entirely inside your own network. Code completion and code search ship as one Rust binary with a web UI, and the repository indexer builds a real retrieval layer over your codebase rather than pasting files into a prompt. Chat and the knowledge base still ship in the v0.32.0 release, but were removed from the main branch by PR #4425 (merged 2026-01-24).

Meng Zhang and Lucy Gao founded TabbyML, Inc.; Meng previously spent four years at Google working on generative AI models. Deployment stays simple: a single node, on GPU or CPU. Note that Tabby collects usage statistics by default; set TABBY_DISABLE_USAGE_COLLECTION=1 to turn that off.

Code completion and code search can run entirely inside your own network from one Rust binary, on GPU or CPU.

Sources

Official

  1. TabbyML/tabby (GitHub)GitHub · Accessed Oct 26, 2026
  2. Tabby READMEGitHub · Accessed Oct 26, 2026
  3. tabby/LICENSE at main · TabbyML/tabbyGitHub · Accessed Oct 26, 2026
  4. Tabby Enterprise License (ee/LICENSE)GitHub · Accessed Oct 26, 2026
  5. Release v0.32.0 · TabbyML/tabbyGitHub
  6. Release v0.31.0 · TabbyML/tabbyGitHub · Accessed Oct 26, 2026
  7. Release next-alpha · TabbyML/tabbyGitHub
  8. feat: sunset knowledge and chat features · Pull Request #4425 · TabbyML/tabby (merged 2026-01-24)GitHub · Accessed Oct 26, 2026

Updates & reading