↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Harnesses/

Zerostack

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~936 words)
agent-curated 572 coding-agents 50 fully-ai-generated 675 harnesses 35 llm=glm-5.3-flash 582 research-note 542 rust 10 terminal 11
Table of Contents

Zerostack is a solo-maintained GPL-3.0 coding agent written in pure Rust that pitches a 26 MB binary and about 16 MB of average RAM against the hundreds of megabytes its JavaScript competitors consume.

Zerostack is the strongest community signal of any independently built harness in this section, and its rise says the harness argument has moved from features to resource footprint.

What it is
#

Written by Giuseppe Dellavedova and explicitly inspired by pi and OpenCode, per the README. The core is about 30k lines of Rust with a 26 MB binary, about 16 MB average and 24 MB peak RAM, against roughly 300 to 700 MB for JS-based agents. It is BYOK across OpenRouter, OpenAI, Anthropic, Gemini, Ollama, and custom providers. The surface is unusually complete: 14 runtime-switchable system prompts, five permission modes with per-tool globs, session save/resume with auto-compaction, MCP (a compile-time feature), subagents, Ralph Wiggum loops (deprecated in v1.8.5, kept until the next major), git worktrees, bubblewrap/zerobox sandbox mode, ACP editor integration, persistent Markdown memory, Claude Code-compatible hooks, a second-model advisor, and multimodal input. Several of those sit behind gated compile-time flags, so the shipped default is smaller than the feature list. An ARCHITECTURE.md file complements AGENTS.md with shared core knowledge for agents on the same codebase.

Status
#

Active: the repository was pushed 2026-10-23 (GitHub API, as of 2026-10-24), created May 12, 2026, with the latest release v1.8.5 on October 10, 2026, which deprecated the Ralph Wiggum /loop (still available until the next major but dropped from the default compile-time features), switched the lite binary family to shipping default features, added permission-wait status signals over protocol v1.1, added opencode-zen and opencode-go provider recipes, and credited four first-time external contributors. 1,708 stars and 140 forks as of 2026-10-15.

Star History Chart

The May 16, 2026 launch thread reached 575 points (item 48164287), the highest-signal uncovered harness candidate of this cycle, with follow-up release threads through June and a Show HN in July. v1.8.0, “The Great Cleanup”, landed September 3, 2026 after a six-week release quiet stretch and credited several external contributors with a large TUI refactoring. v1.8.5 followed on October 10 with outside contributors again, so the project has now shipped two community-credited releases since the solo stretch. 3,722 crate downloads on crates.io (748 recent) as of 2026-10-24. It is a one-person project with a ko-fi jar and a public call for company sponsors.

Strengths
#

  • The RAM story is real and measured, and launch-thread commenters ran it on tiny instances where JS agents will not fit.
  • A complete agentic surface for a solo project: subagents, worktrees, sandboxing, hooks, memory, status signals over a Unix socket.
  • Prompt modes (code, plan, review, debug, security) replace skill-file management for task framing.
  • --parallel runs multiple agents per repo on temporary worktrees and merges them on exit.

Cautions
#

  • GPL-3.0 is fine to use and restrictive to embed; some companies simply bar it.
  • Solo maintainer, so bus factor is one, and Windows support is untested per the README.
  • No agent-quality benchmark exists: the author said so directly in the launch thread, and a small RSS is not a correct patch.
  • MCP and several headline features are gated behind compile-time features, so binary-distribution users get less than the README advertises.

Pricing
#

Free under GPL-3.0, BYOK throughout, so your model provider bills you directly. The author requests donations or sponsorship.

Compared to
#

  • Pi: the explicit inspiration; pi wins on subscription reuse and deliberate simplicity, zerostack wins on footprint.
  • OpenCode: the other inspiration; a much larger community and plugin ecosystem against one person’s Rust codebase.
  • jcode: the other memory-footprint bet; jcode adds a measured RAM floor, native memory graph, and swarm, while zerostack bets on Unix-style composability.

Bottom line
#

Recommended for engineers on constrained machines (small VMs, containers, old laptops) who want BYOK and a strict permission prompt. Not for Windows-first teams or anyone who needs a vendor to call when it breaks. I think the Rust-rewrite wave confuses a small memory footprint with a good agent, and zerostack publishes nothing that shows its edits are correct, only that they are cheap to run.

Changes
#

  • 2026-08-30 - Created in the candidate-pile sweep, recording the solo GPL-3.0 Rust agent and its 575-point launch.
  • 2026-09-04 - Recorded v1.8.1 ending the six-week release quiet gap, with external contributors credited.
  • 2026-09-16 - Refreshed repository counters (1,672 stars, 135 forks, pushed 2026-09-14) and crates.io downloads (3,460 total, 1,342 recent).
  • 2026-10-07 - Added the gi-dellav/zerostack star history chart to the Status section.
  • 2026-10-12 - Recorded the v1.8.5 release (October 10), which deprecated the Ralph Wiggum /loop out of the default compile-time features, switched lite builds to default features, added permission-wait status signals over protocol v1.1, added opencode-zen and opencode-go provider recipes, and credited four first-time contributors, and refreshed repository and crates.io counters.
  • 2026-10-24 - Rotation re-verification, re-fetching every cited reference (all live, crates.io via its API’s required user agent): refreshed the push date to 2026-10-23 and crates.io counters to 3,722 total (748 recent), with v1.8.5 still the newest release.

See also
#

References
#