🧠 Deploy Garry Tan's GStack on Your Ubuntu: Turn Claude Code + Gemini CLI into a Virtual Engineering Team

TL;DR

GStack is Garry Tan's (YC President) personal AI engineering system — 23 opinionated slash commands that turn Claude Code into a virtual team of specialists (CEO, Designer, Eng Manager, QA, Security Officer). You can install it on your Ubuntu VPS alongside your existing Hermes Agent, Claude Code CLI, and Gemini CLI in 30 seconds. It adds structured workflows for product discovery, architecture review, visual design, QA with a real browser, security audits, and release management — all MIT licensed, free, and designed for parallel sprints.

Context

Your Ubuntu system already has:

Hermes Agent instance
Claude Code CLI
Gemini CLI

You're manually prompting these tools. GStack adds opinionated structure — instead of a blank prompt, you get specialized roles that enforce best practices:

Problem	GStack Solution
"Build a feature" → vague, misses edge cases	`/office-hours` + `/autoplan` — 6 forcing questions before code
No design review → AI slop UI	`/plan-design-review` — rates dimensions 0-10, fixes slop
QA is manual → bugs ship	`/qa` — opens real browser, clicks through flows, auto-fixes bugs
Security is an afterthought	`/cso` — OWASP Top 10 + STRIDE, 8/10 confidence gate
No memory across sessions	`/setup-gbrain` — persistent knowledge base (Supabase or local)
Stale docs after shipping	`/document-release` — auto-updates README, ARCHITECTURE, CLAUDE.md

GStack works on 10 AI coding agents — including Hermes (via --host hermes), Codex, Cursor, and OpenClaw.

The Approach

Step 1: Install GStack on Your VPS (30 seconds)

SSH into your Ubuntu VPS, open Claude Code (or any agent), and paste:

git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup

The setup script auto-detects which agents you have installed. For your Hermes Agent specifically:

# After the main clone, install Hermes host support
cd ~/.claude/skills/gstack && ./setup --host hermes

This installs GStack skills to ~/.hermes/skills/gstack-*/.

Step 2: Add GStack to Your Project's CLAUDE.md

From inside your project repository, run:

cd ~/.claude/skills/gstack && ./setup --team
~/.claude/skills/gstack/bin/gstack-team-init required
git add .claude/ CLAUDE.md && git commit -m "require gstack for AI-assisted work"

This adds a ## gstack section to CLAUDE.md with all 23 skills listed and routing rules. Teammates get auto-updates.

Step 3: Run Your First GStack Sprint

# 1. Product discovery (replaces vague prompts)
/office-hours
# Claude asks 6 forcing questions about your idea

# 2. Generate a fully reviewed plan
/autoplan
# Runs CEO → design → eng review automatically

# 3. Build (your existing workflow)
# [implement the plan]

# 4. Review and QA
/review           # catches bugs that pass CI
/qa https://staging.your-app.com  # real browser testing

# 5. Ship
/ship             # syncs main, runs tests, opens PR
/land-and-deploy  # merges, waits for CI, verifies production

# 6. Retrospective
/retro            # weekly engineering retro across all projects

Step 4: Set Up Persistent Memory (Optional but Powerful)

/setup-gbrain

Choose:

Supabase (cloud) — paste a Personal Access Token, auto-provisions a project in ~90 seconds
PGLite (local) — zero accounts, zero network, ~30 seconds

GBrain gives your agents memory across sessions: gbrain search, gbrain put_page, etc. Per-repo trust policies (read-write / read-only / deny).

Step 5: Use GStack with Gemini CLI and Hermes Agent

GStack works across all your agents. Example workflow:

# In Claude Code with GStack
/office-hours  # writes design doc

# Then in Gemini CLI (GStack skills available)
/plan-eng-review  # same structured review

# Then in Hermes Agent
/ship  # consistent release process

For OpenClaw (if you use it), add to AGENTS.md:

## Coding Tasks
When spawning Claude Code sessions for coding work, tell the session to use gstack skills:
- Security audit: "Load gstack. Run /cso"
- QA test a URL: "Load gstack. Run /qa https://..."
- Build a feature: "Load gstack. Run /autoplan, implement, then run /ship"

Why It Worked

1. Structured workflows prevent AI slop

Blank prompts produce inconsistent quality. GStack's forcing functions (/office-hours pushes back on your framing, /plan-design-review rates UI dimensions 0-10) bake in 20 years of product-building experience.

2. Real browser QA catches what tests miss

/qa launches a real Chromium browser (not headless by default), clicks through flows, finds visual bugs, and auto-fixes them with atomic commits + regression tests. Sidebar agent lets you type natural language like "navigate to settings and screenshot it" — the AI executes it live.

3. Parallel sprints scale your output

Conductor runs multiple Claude Code sessions simultaneously — one brainstorming, one reviewing a PR, one implementing, one QAing. Garry Tan reports 10-15 parallel sprints as the practical max. Your VPS can handle this.

4. Cross-model second opinions catch blind spots

/codex runs an independent review from OpenAI's Codex CLI. When both /review (Claude) and /codex (OpenAI) have reviewed the same branch, you get a cross-model analysis showing overlapping and unique findings.

5. Real results: 810× productivity improvement

Garry Tan's measured output (logical code changes normalized for AI inflation) went from 14 lines/day in 2013 to 11,417 lines/day in 2026 — 810× faster. 240× the entire 2013 year already in Q1 2026.

Apply It Yourself

Prerequisites on Your Ubuntu VPS

Ubuntu 20.04+ with Git, Bun v1.0+, Node.js
Claude Code CLI installed and authenticated
(Optional) Gemini CLI, Hermes Agent
API keys for Claude (required), optionally OpenAI/Gemini for /codex

Quick Start Checklist

# 1. Install GStack
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack
cd ~/.claude/skills/gstack && ./setup

# 2. Install Hermes host support
./setup --host hermes

# 3. Add to your project
./setup --team
bin/gstack-team-init required
git add .claude/ CLAUDE.md && git commit -m "add gstack"

# 4. Test a skill
/office-hours
# Describe what you're building

# 5. Run a full sprint on a small feature
/autoplan
# Review the plan
# Implement
/review
/qa http://localhost:3000
/ship

Commands You'll Use Daily

When you're...	Command
Starting something new	`/office-hours` → `/autoplan`
Reviewing code	`/review`
Testing a web app	`/qa <url>`
Shipping	`/ship` → `/land-and-deploy`
Debugging a production issue	`/investigate`
Running a security audit	`/cso`
Updating docs after shipping	`/document-release`
Doing a weekly retro	`/retro`
Exploring designs visually	`/design-shotgun` → `/design-html`
Getting a second opinion	`/codex`
Being extra careful (`sudo`-like)	`/careful` or `/guard`

Troubleshooting Common Issues

Skill not showing up?

cd ~/.claude/skills/gstack && ./setup

/browse or /qa fails?

cd ~/.claude/skills/gstack && bun install && bun run build

Stale install? Run /gstack-upgrade — or set auto_upgrade: true in ~/.gstack/config.yaml.

Hermes skills not loading?

# Verify installation
ls ~/.hermes/skills/gstack*/
# Reinstall if empty
./setup --host hermes --force

Want shorter command names? (/qa not /gstack-qa)

cd ~/.claude/skills/gstack && ./setup --no-prefix

Voice input (bonus for hands-free)

GStack skills have voice-friendly trigger phrases. Say "run a security check", "test the website", or "do an engineering review" — the right skill activates without remembering slash commands.

Uninstall (if needed)

~/.claude/skills/gstack/bin/gstack-uninstall
# Optional: keep config with --keep-state

Source

GitHub Repository: garrytan/gstack
License: MIT — free forever, no premium tier
Documentation: See docs/ folder in repo for skill deep dives, architecture, browser reference, and GBrain setup
Real-world methodology: Garry Tan's productivity data — 810× improvement from 2013 to 2026, 1,237 contributions YTD, 10-15 parallel sprints