Garry Tan's gstack: An Open Source AI Agent Framework for Claude Code

What gstack is
gstack is Garry Tan's open source software factory that transforms Claude Code into a managed virtual engineering team. It's MIT licensed and available on GitHub. According to Tan, he's written over 600,000 lines of production code in the last 60 days (35% tests) and does 10,000-20,000 usable lines of code per day as part of his CEO duties at Y Combinator.
How it works
The framework provides 13 specialist roles as slash commands, all using Markdown. These include:
/plan-ceo-review- Rethinks the product direction (e.g., expands "photo upload" to auto-identify products, pull specs and comps from the web, and draft listings automatically)/plan-design-review- Catches AI slop patterns, provides design scores, exports DESIGN.md, flags patterns like gradient hero, icon grid, uniform radius/plan-eng-review- Locks architecture with ASCII diagrams for data flows, state machines, error paths, plus test matrices and security concerns/review- Finds production bugs, auto-fixes issues like orphan S3 cleanup, missing indexes, and flags race conditions
Other roles mentioned include QA lead (opens real browser and clicks through apps), release engineer (ships PR), and designer.
Example workflow
From the source: When requesting "photo upload for sellers," /plan-ceo-review responds with expansion proposals, /plan-design-review provides an 80-item audit and design system inference, /plan-eng-review creates ASCII diagrams and test matrices. After approval, Claude writes 2,400 lines across 11 files (models, services, controllers, views, migrations, tests) in about 8 minutes.
Technical details
The repository includes directories for workflows, bin, design-consultation, docs, document-release, gstack-upgrade, plan reviews, QA reviews, retro, review, scripts, setup-browser-cookies, ship, and test. Key files include ARCHITECTURE.md, BROWSER.md, CLAUDE.md, conductor.json, and package.json.
📖 Read the full source: HN AI Agents
👀 See Also

Claude Sessions: Lightweight Desktop App for Browsing Claude Code History
Claude Sessions is a new desktop application that lets developers browse their Claude Code session history locally. It reads from ~/.claude/projects, organizes sessions by project, handles large sessions up to 500k+ tokens without lag, and includes search functionality and keyboard navigation.
DashClaw: An Approval Layer for OpenClaw Agents
DashClaw is an open-source approval layer that intercepts destructive tool calls in OpenClaw, Claude Code, and other agents, requiring human approval via Telegram before execution.

PageAgent: Browser AI Agent That Runs Inside Web Pages with Ollama Support
PageAgent is a JavaScript library that runs AI agents directly inside web pages, reading live DOM as text instead of using screenshots. It works with any OpenAI-compatible endpoint including Ollama, enabling local LLM calls directly from the browser.

Codiff v0.1.0: A Local Diff Viewer for LLM-Generated Code Reviews
Codiff v0.1.0 is a fast, minimal desktop app for reviewing local Git diffs, with LLM walkthrough mode and inline comments that can be copied as Markdown.