Free
$0
Open-source coding agent under GPL-3.0. Install via cargo or pre-built binaries from GitHub Releases. No paid tiers.
Zerostack
A Unix-inspired, radically minimalistic coding agent written in pure Rust — ~16MB RAM, 12.9MB binary, 12k lines of code.
Zerostack is a Unix-inspired, radically minimalistic coding agent written in pure Rust — ~16MB RAM, 12.9MB binary, 12k lines of code. It supports OpenRouter, OpenAI, Anthropic, Gemini, and Ollama providers, includes a crossterm-based TUI, MCP support, an integrated loop system for long-horizon tasks, Git worktree integration, and an ACP (Agent Communication Protocol) server for editor integration.
+Pros
- Radically lightweight (~16MB RAM vs 300-700MB for Claude Code/OpenCode)
- Pure Rust — no Node/Go/Python runtime dependency
- Integrated loop system for long-running autonomous tasks
- Git worktree integration for branch-per-task workflows
- ACP server enables Zed editor integration
- Multi-provider support including Ollama for local models
- 10+ built-in prompt modes (code, plan, review, debug, ask)
- Granular permission system with doom-loop detection
−Cons
- Very early stage (less than 2 weeks old, rapid API changes)
- Small community and plugin ecosystem
- TUI-only — no IDE plugin or web UI
- Documentation still maturing
- Windows support untested
- Loop and worktree features marked experimental
Pricing
Introduction
Zerostack is a Unix-inspired, radically minimalistic coding agent written in pure Rust. Its single purpose: be the lightest coding agent you can run — ~16MB RAM at idle, ~12.9MB on disk, CPU at 0.0% when doing nothing.
The coding agent landscape has grown bloated. Claude Code burns 300-700MB of RAM. OpenCode ships a full Node.js runtime. In a landscape of ever-growing resource demands, Zerostack went the other way: pure Rust, ~12k lines of code, a single statically-linked binary that runs on a Raspberry Pi 4.
Launched on GitHub on May 16, 2026, Zerostack hit nearly 1,000 stars within two weeks — a signal that the developer community has been waiting for a lightweight alternative.
At its core, it distills the coding agent to essentials: a crossterm-based TUI, a multi-provider LLM layer (OpenRouter, OpenAI, Anthropic, Gemini, Ollama), a session manager, and a permission system. Then it adds ambitious features — an integrated loop system for multi-step autonomous tasks, Git worktree integration for branch-per-task workflows, and an ACP server that opens the door to editor integration with tools like Zed.
Key Features
- Multi-Provider Support: Connect to OpenRouter, OpenAI, Anthropic, Google Gemini, Ollama, or any custom OpenAI/Anthropic-compatible endpoint. Provider configuration lives in a single JSON file.
- Tool Ecosystem: Read/write files, run shell commands, execute lua scripts, search with ripgrep, use semantic search (via Exa), and control the agent via tools. Tools are structured as a typed JSON function API — new tools can be added by dropping a Rust file into the tools directory.
- Monochrome Mode: Terminal-native TUI using crossterm, with a Monochrome mode available for minimal visual interference. Runs in any terminal emulator.
- Built-in Prompts: 10+ prompt modes (code, plan, review, debug, ask, brainstorm, and more) that template the system prompt and initial user message for common workflows.
- Permission System: Four permission modes (accept-all, accept-per-tool, per-tool-glob, ask-each), per-tool patterns, and doom-loop detection that halts tool calls when recursion is detected.
- Session Management: Persistent session history backed by a configurable data directory. Sessions resume across restarts. Each session tracks tool calls, token usage, and file modifications.
- Loop System (experimental): The loop system handles multi-step autonomous tasks by maintaining a plan, tracking completion, and iterating until done. Activate it with --loop, and Zerostack will plan, execute, validate, and refine until the task is complete — optionally creating a new Git branch for each task.
- Git Worktrees (experimental): Create isolated Git worktrees for each task, keeping your main working directory clean. This works for experiments, code reviews, and long-running feature branches without polluting the working tree.
Version History
Release v1.3.3
Release v1.3.2
Release v1.3.0
Initial public release
Signature Snippet
# Install via cargo
cargo install zerostack
# Launch
cd /my-project
zerostack
# Run with loop for long tasks
zerostack --loop
# Create a task in a separate branch
git worktree add -b feature/validate-api ../zt-validate-api
cd ../zt-validate-api && zerostack --loop "Add validation to the user API endpoint"More in this Space
Prime Agent (Prime Intellect)
Prime Agent is an open-source (MIT) self-improving coding and research harness by Prime Intellect, built on a Recursive Language Model that runs inside a persistent IPython REPL and a Continual Harness that refines its own state via /refine. It took the same frontier weights from 30.2% to 95.5% on ARC-AGI-3 RHAE Best@1, powered daemon-backed sessions, and ships no hosted offering.
OpenCodeReview (ocr)
OpenCodeReview (ocr) is Alibaba's open-source AI code-review CLI. It reads git diffs and drives an LLM agent through a hybrid architecture where deterministic pipelines handle file selection, rule matching, and comment positioning while the agent does the reasoning. Built on two years of internal use at Alibaba scale, it ships a multi-language ruleset and consumes roughly 1/9 of the tokens of general-purpose agents.
Delta (by Zed)
Multiplayer environment from Zed for coding with agents. Threads unify the agent conversation, the code edits, and human review in one shareable artifact, replacing the branch/PR loop. Built on DeltaDB, a CRDT-based version control layer that records every edit between commits alongside the prompts and reasoning that produced them.