Builders Spotlight — Tabby

Tabby

A self-hosted GitHub Copilot alternative built by TabbyML that runs code completion models locally with IDE integration and zero telemetry.

The problem it set out to solve

Developers wanted AI-assisted coding without sending their code to external servers or paying per-seat licenses. GitHub Copilot worked well but created a privacy tradeoff — your codebase became training data fodder for a cloud service. For teams with proprietary code, regulated industries, or just privacy concerns, the choice felt binary: use Copilot and accept the risk, or go without. Tabby exists because that binary shouldn’t have existed.

The key insight

Code completion is a solved problem in terms of model capability — you don’t need a frontier model like Claude Opus 5 or GPT-5.6 Sol to suggest the next line. What you need is a system architecture designed specifically for the IDE integration layer and the inference constraints of real-time completion. Tabby decoupled the model serving infrastructure from the IDE plugin, building a server-first approach where any editor can connect to a self-hosted completion backend. The philosophy: make it as easy to self-host a code model as it is to run a Postgres database.

How it works (in plain terms)

Tabby runs as a daemon that listens for completion requests from IDE extensions (VS Code, JetBrains, Vim, etc.). When you stop typing, your editor sends local context to the Tabby server, which uses a specialized code model to generate completions. The server handles batching and caching to make inference fast enough for interactive use. Everything stays on your infrastructure — no cloud calls, no telemetry. The trade-off is intentional: you run the server yourself, which adds operational overhead, but you keep full control of your data and can customize the model for your codebase.

What it looks like in practice

# Start Tabby server
tabby serve --model StarCoder-1B

# Connect from VS Code via extension settings
# Set "Tabby: Server Endpoint" to http://localhost:8080

# Now trigger completion (Ctrl+\ in VS Code)
# Tabby sends your file context and returns suggestions in real-time

The server exposes a simple HTTP API that IDE plugins consume. No authentication setup required for local networks.

Why it matters

• Privacy as default: Code never leaves your infrastructure. For teams handling financial, medical, or classified systems, this isn’t a nice-to-have — it’s mandatory.
• Cost predictability: No per-seat licensing or usage-based billing. You pay once (in compute), not continuously.
• Customization opportunity: Self-hosting means you can fine-tune Tabby’s models on your codebase patterns, making suggestions more relevant to your specific tech stack and conventions.

Where to go next

• Tabby GitHub repository — source code, deployment guides, and contributing guidelines
• Tabby docs — setup instructions for different editors and deployment targets (Docker, Kubernetes, bare metal)
• TabbyML blog — write-ups on model selection, performance tuning, and community-contributed integrations