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
Local MCP Memory System with Consolidation for AI Conversations
A developer built an MCP server that provides persistent local memory for AI clients, using Qwen 2.5-7B to consolidate conversations into structured knowledge documents every 6 hours. The system runs entirely on your hardware with semantic dedup, adaptive scoring, and FAISS vector search.
Layerkit: AI Image Editor with Editable Layers Built with Claude Code
A developer built Layerkit, a browser-based AI image editor that generates scenes with editable layers to avoid constant re-prompting. The tool uses a multi-stage AI pipeline where one LLM plans composition, an image model generates the scene, and another LLM analyzes the actual image to place readable text.
OpenClaw Codex-GPT5.4 Task Validation Loop Issue
A developer reports Codex-GPT5.4 through OpenClaw gets stuck in a task validation loop during autonomous project work, repeatedly identifying and confirming tasks without executing them. They've implemented workspace controls including TASKS.md, heartbeat rules, and persona files to address the issue.
Custom Status Line for Claude Code Shows Context Usage, Cost, and Git Branch
A Reddit user created a bash script that leverages Claude Code's statusLine setting to display real-time information including context window usage, session cost, active model, and current git branch. The script requires jq and is available on GitHub.