Snowtree: Databend's Best Practices for AI-Native Development

Why Snowtree?

Databend is a modern, open-source cloud data warehouse built in Rust. It's a massive system - over 1.9 million lines of code across 11,000+ files - powering critical Analytics, Search, and AI workloads. For us, correctness, performance, and stability aren't just features; they are non-negotiable.

When we shifted to an AI-first development culture, we evaluated many mainstream AI coding tools. Unfortunately, most fell short of the rigor required for serious systems engineering:

• Review Fatigue & Technical Debt: Complex systems have subtle, implicit constraints that AI often misses. When an agent generates massive changes, reviewing them becomes a nightmare. In multi-turn sessions, an agent can often clobber code you just approved, forcing you to review the same logic repeatedly. Without a strict, incremental review process, AI-generated code quickly turns into unmaintainable technical debt.
• Fragmented Workflows: With everyone picking their favorite model or IDE plugin, establishing a consistent team standard or development flow became impossible.
• The "Wrapper" Gap: Many AI IDEs wrap models in their own custom agents. This seemingly convenient layer often blocks access to the model's native capabilities - like Claude's MCP or Skills - and new features arrive only when the IDE updates, not when the model does.

We needed a tool that unleashes AI's potential without surrendering human control. A tool that standardizes our workflow, preserves native model capabilities, and integrates seamlessly with GitHub.

So we built Snowtree. It's not just a tool; it's the codified workflow of how Databend builds database internals with AI.

The "Snowtree" Philosophy

The name is a nod to Git's worktree feature.

Before AI, we rarely touched worktrees;

git branch

This matches perfectly with worktrees: native isolation. Every AI task gets its own pristine working directory (

worktree

Snowtree manages these ephemeral environments. Like snowflakes, each workspace is distinct and temporary, eventually merging seamlessly into the main branch.

Snow + Worktree = Snowtree.

The Workflow: AI Speed, Human Safety

1. Isolate

Snowtree spins up a sandbox for your agent. The AI can refactor, break things, and experiment freely without polluting your main checkout or staging area.

2. Review

This is where the magic happens. When the AI proposes changes, Snowtree presents a clean, interactive diff. You review line-by-line: Keep what's solid, Discard what's hallucinated or sloppy. The AI fixes only what you rejected.

3. Iterate

Every review acts as a checkpoint. Once you approve a chunk of code, it's staged and safe - subsequent AI turns won't accidentally overwrite it. When the task is done, you commit and sync a Pull Request in one go.

No Middleman: Pure Native CLI

Snowtree is a workflow orchestrator, not a wrapper. It invokes the official CLIs - Claude Code, Codex, Gemini CLI - directly.

This means zero capability loss. If you configure a Databend MCP server in Claude Code's Model Context Protocol (MCP), it works in Snowtree instantly. When Claude Code or Codex releases a new feature, you get it the second you update their CLI, without waiting for us to "support" it.

Quick Start

Install Snowtree:

curl -fsSL https://raw.githubusercontent.com/databendlabs/snowtree/main/install.sh | sh

Install Your Agent of Choice:

npm install -g @anthropic-ai/claude-code

npm install -g @openai/codex

npm install -g @google/gemini-cli

Is Snowtree for You?

Snowtree isn't for everyone. But it might be exactly what you need if:

• You are building production-critical infrastructure where "looks complete" isn't good enough.
• You work in a team and rely on GitHub for collaboration.
• You manage a complex codebase and need a systematic way to review high-volume AI contributions.

Acknowledgements

Snowtree stands on the shoulders of giants. We drew deep inspiration from Zed's fluid diff interactions and OpenCode's minimalist TUI philosophy. These tools showed us that dev tools can be both powerful and distraction-free.

• GitHub: databendlabs/snowtree