Garry Tan Releases gstack: An Open-Source Workflow Layer for Claude Code
These articles are AI-generated summaries. Please check the original sources for full details.
Eight opinionated workflow skills for Claude Code
Garry Tan has released gstack, an open-source toolkit designed to add structured operational boundaries to Claude Code. The system introduces a persistent browser daemon that reduces tool call latency from 5 seconds down to approximately 200ms.
Why This Matters
In typical agentic workflows, browser automation often suffers from high latency and state loss because fresh instances are launched for every action. gstack addresses this by running a long-lived headless Chromium process that maintains cookies, localStorage, and tabs across commands, enabling a more reliable bridge between LLM reasoning and application state. By enforcing explicit modes for planning, review, and QA, the system moves away from monolithic agent prompts toward specialized, role-bound execution. This structured approach ensures that product-level planning and engineering risk assessment remain distinct phases, reducing the likelihood of architectural drift during automated shipping.
Key Insights
- gstack version 0.3.3 uses a long-lived Chromium daemon over localhost HTTP to maintain browser state across commands.
- Persistent sessions reduce tool latency from 3–5 seconds to roughly 100–200ms by eliminating cold starts.
- The /qa skill maps branch diffs to affected application routes, enabling targeted testing of specific UI flows.
- The toolkit is built on Bun v1.0+ to utilize native SQLite access for reading Chromium cookie databases and compiling native binaries for macOS and Linux.
- Operational structure is enforced through 8 core commands including /plan-ceo-review, /plan-eng-review, /ship, and /retro.
Practical Applications
- Automated QA: gstack analyzes 8 changed files and 3 affected routes to test local application instances automatically. Pitfall: Treating QA as a detached manual pass rather than a diff-driven automation step.
- Stateful Browser Automation: Using /setup-browser-cookies to import local browser sessions into headless Chromium for authenticated agent actions. Pitfall: Managing separate runtime toolchains within skill directories instead of using compiled binaries like those provided by Bun.
References:
Continue reading
Next article
Integrating Strands Agents with Amazon Bedrock AgentCore Browser for Autonomous Web Tasks
Related Content
Nous Research Unveils Hermes Agent: Solving LLM Forgetfulness with Multi-Level Memory and Persistent Terminal Access
Nous Research releases Hermes Agent, an open-source system featuring a multi-level memory hierarchy and persistent machine access to eliminate AI state decay.
Alibaba Releases OpenSandbox: A Unified, Secure API for Autonomous AI Agent Execution
Alibaba launched OpenSandbox under the Apache 2.0 license, providing a unified API for AI agents to securely execute code and browse the web in isolated environments.
Microsoft Releases Agent Lightning: A Reinforcement Learning Framework for Optimizing AI Agents
Microsoft introduces Agent Lightning, an open-source framework that enables reinforcement learning (RL)-based training of large language models (LLMs) for AI agents without requiring changes to existing agent stacks.