tabby.ml

Paid

tabby.ml is an open-source, self-hosted AI coding assistant for developer tools. You run it on your own hardware. Your code never leaves your machine.

You get code completion, an answer engine for coding questions, and inline chat. It plugs into VS Code, IntelliJ, and niche JetBrains tools like AppCode and CLion.

It fits teams with strict privacy rules or air-gapped networks. It is a poor fit if you want one-click setup.

Open it if you need private, self-hosted code completion. Skip it if you want managed cloud convenience.

It helps most when data control matters more than easy setup.

• Teams with strict data residency requirements: You run everything yourself. No cloud provider sees your code.
• Developers using JetBrains IDEs beyond the main three: You get support for AppCode, CLion, and Android Studio, not just IntelliJ.
• Organizations wanting supply chain transparency: You can audit the Apache 2.0 source. No black-box dependencies.
• Developers comfortable with Docker and Rust build tooling: You set up containers and manage the build yourself.
• Users with consumer-grade GPUs: Tested on RTX A5000 and Apple Silicon. You do not need enterprise hardware.
• Teams wanting to avoid Copilot's subscription and cloud terms: You own the stack. No monthly fees for small teams.
• Bitbucket users wanting pipeline automation: You can wire it into pull request workflows with custom scripts.
• Security-conscious teams who prioritize data residency: You keep full audit logs. No third-party inference.

These are the parts you notice after running it for a week.

• Your code stays private: Nothing leaves your network. This matters for regulated industries and proprietary codebases.
• Repository-aware context: It pulls your actual project structure and imports into suggestions. You get completions that match your patterns.
• Lower latency than cloud tools: Local inference hits ~30ms versus ~200ms for remote calls. You feel the speed on every keystroke.
• Multiple inference backends: You can switch between llama.cpp, vLLM, and TensorRT-LLM based on your hardware.
• Free for small teams: Up to 5 users with no feature gates. You only pay for your own GPU and time.

The path is not short, but it is clear.

1. Install Docker and prepare your GPU drivers.
2. Run the container with your chosen model and hardware flag.
3. Install the IDE extension and point it at your local server.
4. Open a file and start typing to trigger completion.

Skip it when the overhead outweighs the benefit.

• No managed option: You handle uptime, updates, and GPU drivers yourself. There is no support line to call.
• Setup complexity is real: Docker, Rust toolchain, CUDA or Metal setup. Plan for an afternoon, not a minute.
• Model quality is capped: Local models like StarCoder-1B are smaller than cloud frontier models. Complex tasks may disappoint.
• Documentation is scattered: You chase links across repos and external sites. Expect to dig for answers.