GStack Engineering Workflow
Turn Claude Code into a role-based engineering workflow for planning, review, QA, and releases.
- Source repo
- garrytan/gstack
- Stars
- ★ 137k
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 80/100 · Good
At a glance
- How it runs
- Works with
- Portable with changesClaude CodeCodex (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A technical founder with a vague product idea can start with
/office-hoursand let later planning skills consume its design document before implementation. - Not a fit if
- Teams seeking a fully hosted service
- Developers unwilling to set up Claude Code, Git, and Bun
- Source review
- 80/100 · Good
What does this agent do, and when should you use it?
gstack is a collection of Markdown skills and standalone command-line tools for Claude Code, connecting product planning, engineering design, code review, QA, and release work. After installation, slash commands such as /office-hours, /plan-ceo-review, /review, /qa, and /ship start role-based workflows; earlier steps write design documents and test plans for later skills to use. Browser tasks use Aside first, and fall back to the Chromium browser built during setup when Aside is absent or closed. The repository also documents installation for hosts including Codex CLI, OpenCode, and Cursor, though Claude Code is the default and the only host marked fully certified, while several others remain experimental. It runs locally with shell, filesystem, and network access and calls configured AI CLIs or APIs as workflows require; it is not a hosted SaaS.
The ./setup installer can add skills to Claude Code or select other documented hosts with --host. /office-hours questions product requirements and produces a design document; /plan-ceo-review and /plan-eng-review review product scope and engineering design; /review inspects changes and can fix obvious issues; /qa explores browser, API, CLI, job, or webhook behavior and, when available, writes regression tests, fixes defects, and verifies them; /ship runs verification, reviews coverage and documentation, pushes changes, and opens a PR. /land-and-deploy can merge an approved PR and check deployment, while /cso performs a static security assessment. Additional skills and standalone CLIs support design, documentation, diagrams, PDFs, browser work, and workflow utilities. Outputs include design documents, plans, review and QA evidence, tests, documentation changes, and PRs; the exact operations depend on the chosen skill and its required tools.
- A technical founder with a vague product idea can start with
/office-hoursand let later planning skills consume its design document before implementation. - An engineering lead preparing a PR can run
/reviewon a changed branch, then use/shipfor release checks. - A developer validating a staging site or local API, CLI, job, or webhook fixture can use
/qato explore behavior and preserve regressions when the setup supports it. - A product design team comparing UI directions can generate mockups with
/design-shotgunand turn a selected direction into HTML with/design-html. - A team using several AI coding hosts can install for a documented experimental host, or run a focused skill for security assessment or release preparation.
How do you install or deploy this agent?
Install Claude Code, Git, and Bun (the documented tested version is v1.4.2+; setup refuses versions below v1.3.3). Clone the repository and run setup in a terminal:
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack
cd ~/.claude/skills/gstack
./setupWindows also requires Node.js. To install for another documented host, use --host, for example Codex CLI:
./setup --host codexFor the shared-repository team mode, the documented example enables team setup and initializes repository configuration:
(cd ~/.claude/skills/gstack && ./setup --team) && ~/.claude/skills/gstack/bin/gstack-team-init requiredHow do you use this agent?
In Claude Code, start with a planning or review skill. One documented quick path is:
/office-hours
/plan-ceo-review
/review
/qa https://staging.example.com
/ship/qa can also target a local CLI or API fixture; describe the target and behavior to check. To let /ship open a PR or /land-and-deploy affect production, configure the project's relevant test, Git, and deployment tools; the latter follows an approval step. Outside reviews require the corresponding CLI to be installed and authenticated. Browser workflows can use Aside or the Chromium browser built by setup. Some design features use OpenAI models by default and need access to the relevant API; requirements vary by command.
What are this agent's strengths and limitations?
- The workflow links
/office-hours, planning reviews,/review,/qa, and/ship, with earlier documents available to downstream skills. /qacovers browser and non-web targets such as APIs, CLIs, jobs, and webhooks, and emphasizes reproducible bugs and regression verification.- Besides Claude Code, setup lists installation paths for hosts such as Codex CLI, OpenCode, and Cursor, and the repository includes standalone CLIs.
- Browser workflows use Aside when available and can fall back to bundled Chromium.
- Installation requires Claude Code, Git, Bun, and a local shell; Windows also requires Node.js, so setup is not zero-config.
- Claude Code is the only host marked full; Codex CLI, OpenCode, Cursor, and several others are marked experimental.
- Outside reviews depend on another installed and authenticated CLI; some design features also depend on OpenAI models or API access.
- Some capabilities have additional platform or tool requirements: iOS QA needs a Mac and a connected iPhone or iPad, while the documented comprehensive CSO mode has no published qualified runtime or scanner profiles yet.
How does this agent compare with similar options?
The README directly contrasts gstack's browser-control approach with browser-use/browser-harness-js: gstack provides curated commands and security controls, while browser-harness-js offers thinner CDP transport for agent-authored helpers. It also says gstack can share design rules and DESIGN.md with impeccable.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| GStack Engineering Workflow This agent | 80 · Good | Agent plugin / skillFree + model costs | ★ 137k | today | TypeScript | Claude Code |
| Itqan Engineering | 80 · Good | Agent plugin / skillFree + model costs | ★ 12 | 21d ago | Shell | Codex · Claude Code |
| Pilot Shell | 53 · Major gaps | CLIPaid | ★ 2.1k | today | JavaScript | Claude Code |
| Maestro Orchestration Platform | 75 · Some gaps | Agent plugin / skillFree + model costs | ★ 486 | today | JavaScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The evidence shows some limits and confirmation gates: review may auto-fix obvious issues but asks approval for other fixes; outside reviews can be disabled and read the setting before sending. Installer documentation says it will not reset or forcibly move dirty, diverged, or pinned checkouts, and upgrades can refuse while active leases exist. Outbound content, model selection, and sources have notices and diagnostics; the guard log is described as content-free. Deductions: repository-level tools can edit code, operate a browser, and publish PRs, so permissions are not minimal; outside reviews are enabled by default, and a first-use notice is transparent but not per-send consent; sensitive-data handling is not established for every tool or storage path; the MIT license and copyright are clear, but publisher identity is unverified and maintenance responsibility is not explicitly committed.
The docs connect planning, review, QA, and release in a stated sequence, distinguish outside-review gaps from passes, and describe install verdicts recording revision, runtime, and scope. Dependencies are named, downloaded runtimes are SHA-256 checked, and browser installation is lazy. Deductions: the README describes broad behavior that cannot all be independently checked against the supplied materials; the docs also acknowledge a residual lease-reclamation race under consecutive process kills, so full marks are not justified.
The docs identify founders, first-time Claude Code users, and technical leads as audiences; list supported hosts, maturity tiers, workflows, and restrictions; and mostly use explicit slash commands with natural-language examples. Deduction: some examples present idealized automatic behavior, leaving the boundary between conversational requests and commands less precise than a full score warrants.
The repository material has a clear install sequence, quick start, host matrix, workflow tables, troubleshooting, stated limitations, and MIT license; it also documents multi-host installation, repair paths, and version information. Deductions: a rename is mentioned, but the supplied material does not provide a complete change history; Garry Tan is named as copyright holder, but no explicit maintenance response or responsibility commitment is given.
Command outputs, gate outcomes, and troubleshooting messages include status, reasons, and repairs; the workflow connects design, planning, review, QA, and release, with plausible coordination savings. Deductions: productivity claims are primarily README assertions and are not independently validated by the supplied material; installation, browser, and outside-model use add costs, and benefits depend on the project and workflow.
Several implementation claims point to specific docs, configuration, or diagnostic entry points, and troubleshooting distinguishes not-run, unavailable, and pass outcomes. Deductions: the supplied evidence is mainly the README and two operational documents, without corresponding code, tests, or independent corroboration; productivity figures and speed examples are author claims with a methodology-document pointer but cannot be checked from the provided material.
- Outside reviews are enabled by default; check the setting and first-use notice before sending code or plans, and disable them if desired.
- Installation and upgrades fetch code, runtimes, or a browser over the network; review target paths, network behavior, and each host's safety tier before deployment.
- Productivity and workflow examples are documentation claims, not independently verified results.
FAQ
Does using gstack cost money?
Can I use it with Codex CLI?
./setup --host codex. The repository marks Codex CLI experimental and says its safety skills are advisory rather than enforced by hooks.Does browser automation require Aside?
Will QA change my code automatically?
/qa is designed to reproduce issues, write regression tests, fix the cause, and re-verify. /qa-only explores and reports without changing product code or tests. Available actions depend on the target environment and permissions.Can it run a full security scan?
/cso can perform static assessment, but no matching qualified runtime images or scanner profiles have been published yet, so the described audits run static assessment.