Zerostack is popular: a Unix-style programming agent written in pure Rust with only 12MB of memory

Core conclusion

On May 17, 2026, an open source project called Zerostack topped the Hacker News hot list (171 points, 46 comments), becoming the most talked about AI programming tool topic of the day.

Three core highlights:

• Extremely lightweight: A blank session is only 8MB, and a working session is 12MB. In comparison, Claude Code can easily take up several GB.
• Unix philosophy: Pure Rust implementation, emphasizing pipes, redirections, least privilege, and scripted interaction modes
• Strong resonance from the community: Many developers mentioned the "memory leak" pain point of Claude Code and OpenCode, and Zerostack is playing right into it.

Background: The debate between "fat" and "thin" programming agents

The current mainstream AI programming agents - Claude Code, OpenAI Codex, OpenCode - are all taking the "big and comprehensive" route: built-in TUI, file system monitoring, context caching, MCP integration, etc. The result is a memory usage spike.

One HN comment summed up this pain point:

"OpenCode slowly leaked memory, eventually swelling to 6GB, and getting slower and slower. Seeing Zerostack is too timely." - 360MustangScope

Another comment puts the problem more directly:

"RAM usage: ~8MB for empty sessions, ~12MB for work. I like this a lot, Claude Code uses several GB, which is really annoying on a low-end laptop." - throwa356262

At the same time, competing products such as Pi.dev and Codex CLI also have their own characteristics. Pi emphasizes the "self-mutation" of models and tool generation capabilities, while Codex relies on OpenAI's model ecosystem. Zerostack's choice is a completely different direction - a return to the simplicity of Unix design.

Key impact: Zerostack’s differentiated positioning

Adaptation suggestions

For AI tool users, the emergence of Zerostack deserves attention:

1. New projects/light scenarios are preferred for trial: When developing a small script or prototype, use Zerostack instead of Claude Code to avoid being slowed down by RAM.
2. Arrangement and Loop: Zerostack's built-in Ralph Wiggum loop is very practical in long task scenarios. It can be understood as "after adjusting the interface once, it will automatically repeat until it is completed."
3. Pipeline combination: Zerostack supports standard Unix pipelines, which can be plugged into existing automated workflows - such as working with n8n to do code review pipelines
4. Security: Community members have used DeepSeek V4 Flash to scan the Zerostack code base, and the conclusion is that "no threats were found"

Actual usage examples

# Zerostack
cargo install zerostack

#
zerostack

# （）
echo " Python HTTP " | zerostack

# （）
zerostack --approval

Related extended information

• HN — 46 reviews with tons of side-by-side reviews and experiences
• crates.io — Pure Rust implementation, no additional runtime dependencies
• Pi.dev — Another programming agent that takes the "streamlined route" (mentioned by HN in the same discussion)
• airun — cross-platform proxy runtime, compatible with multiple prompt formats

Tool entry

Tools featured in this article include Rust, Claude Code, OpenAI Codex, OpenCode, DeepSeek, n8n, Pi.dev.

Internal link guidance

• Want to build your own AI programming workflow? Watch: Claude Code automated writing practice: build an AI content production pipeline in 30 minutes
• Real case: Someone used Claude Code + AWS to build AI SaaS, He used Claude Code + AWS to build AI SaaS, and his monthly income was $12,000 for 3 months
• Want to use automation tools to connect everything? See: How to build an AI content automated distribution system with n8n + ChatGPT: a complete 30-minute tutorial