YC chief's Claude Code toolkit: browser daemon or bloat?

You know that feeling when you ask an AI coding agent to build something and it hallucinates a plan, skips the review step, and ships broken code? Your agent has no structured workflow — it jumps from idea to code with no plan review, no QA, and no reflection on what went wrong. You end up hand-holding the AI through each phase, re-prompting it to test in a browser, and manually checking its work because there's no pipeline enforcing discipline.

Think of gstack as a pre-built team of specialists you drop into your AI coding agent. You install 23 Markdown skill files that Claude Code discovers as slash-commands (`/office-hours`, `/plan-ceo-review`, `/build`, `/review`, `/qa`, `/ship`). The clever part is a persistent Chromium browser daemon — a Bun-compiled TypeScript binary (~58MB) that stays running in the background. When a skill needs to check a webpage, it sends a command to localhost, which drives Chromium via Playwright's accessibility-tree locators (not DOM injection). First call takes ~3 seconds (cold start), subsequent calls ~100-200ms. Each skill's output feeds into the next in a structured pipeline: you describe an idea in `/office-hours`, get a CEO-level plan review, hand off to engineering, run browser-based QA, then ship.

• Persistent Chromium daemon — why you care: browser commands respond in ~100-200ms after first launch, letting your AI agent do visual QA and design review without you manually opening browsers or taking screenshots.
• 31 structured skills (23 specialist + 8 power tools) — why you care: you get a Think-Plan-Build-Review-Test-Ship-Reflect pipeline out of the box instead of ad-hoc prompting; each skill's output feeds into the next.
• Ref-based element addressing via Playwright Locators — why you care: avoids CSP conflicts, framework hydration issues, and Shadow DOM problems that break traditional DOM-based browser automation.
• 8-agent support (Claude Code, Codex CLI, Cursor, OpenCode, Factory, Slate, Kiro, OpenClaw) — why you care: you aren't locked to one AI coding assistant; a declarative host config system adapts skills to each agent.
• Community security auditing (3 waves, 30+ fixes, 7+ external contributors) — why you care: the project has real security scrutiny with Bearer token auth, localhost-only binding, and user-approved Keychain access for coo...
• Atomic state storage (~/.gstack/browse.json, mode 0o600) — why you care: your browser session state is persisted safely between commands, so the daemon picks up where it left off after a crash recovery.

Who It Is For

If you're a developer already using Claude Code (or another AI coding agent) who wants structured workflow discipline and browser-based QA without switching tools. Not useful if you're on Windows and need cookie features, or if your team needs multi-user collaboration (single-user architecture only). Also not for you if you're sensitive to token costs — the full pipeline loads substantial SKILL.m...

Worth Exploring?

Worth a look if you already use Claude Code and want browser-based QA — that's the feature actual users praise most. Install it, try `/office-hours` on a real project, and see if the pipeline discipline sticks. Know what you're signing up for: this is beta software shipping at extreme velocity (multiple versions per day) with no formal releases, macOS-only cookie decryption, and significant token costs on the planning skills. The 72K stars reflect Tan's platform, not necessarily production maturity.