mrkeyoor.com_
Wed 12 Aug 17:46 UTC
LLM Toolsevaluationupdated 12 Aug 2026

gstack

gstack is an opinionated set of AI-agent workflows for product discovery, design, engineering review, browser QA, security, release work, and retrospectives. It began around Claude Code and now installs host-specific skills for Codex, Cursor, OpenCode, and several other coding agents.

Verdict

gstack contains more concrete engineering practice than most viral prompt collections, and its browser, QA, review, and release chain can be genuinely useful. It also grants an agent a wide operational surface and is changing fast enough to produce costly host-specific regressions. Adopt individual skills first, measure token and tool behavior, and earn your way toward the full workflow rather than installing team mode on faith.

Setup3/5One setup script hides substantial host and browser integration
Docs5/5Extensive guides, examples, caveats, privacy notes, and uninstall steps
Community5/5127,674 stars and exceptionally active same-day issue work
Maturity3/5Broad tested surface, but serious fresh regressions remain

Who it’s for

Technical founders who want one prescribed process from idea review through deployment
Claude Code users who prefer explicit specialist roles over an empty prompt
Experienced engineers able to supervise automated edits, browser actions, commits, pushes, and releases
Teams willing to standardize their agent workflow and update it frequently

Who it’s NOT for

Developers who want a few independent prompts, because gstack intentionally becomes an orchestration layer across the whole sprint
Cost-sensitive Codex users until issue 2519 is resolved, because one reported review consumed at least 15.2 million input tokens
Teams that forbid silent tool updates, because recommended team mode checks for and applies updates during sessions
Windows users needing predictable parity, because current reports cover broken hooks, sync scripts, and host-specific runtime paths

Setup reality

Cloning and running setup is fast when Git, Bun, and a supported agent are already installed. Real adoption takes longer: gstack modifies personal or repository skill directories, adds workflow instructions, generates host variants, installs browser support, may configure hooks and telemetry, and can wire auto-updates. Test it outside a valuable repository, inspect the generated files, and enable team mode only after agreeing on update and release permissions.

An operating system for agent work

gstack is easiest to understand as a prescribed software-making process. Its skills move from product interrogation to executive, design, engineering, and developer-experience review, then into implementation, code review, browser QA, security assessment, release, deployment monitoring, documentation, and retrospectives. Outputs pass between stages, so the agent is expected to remember the original problem and acceptance criteria rather than treat every command as a new chat.

That structure is the real product. The instructions force useful behaviors: expose assumptions, draw data flows, consider failure modes, inspect real diffs, test in a browser, generate regression coverage, and verify deployment health. A technical founder who tends to jump from idea to code can benefit from the friction. A new Claude Code user gets starting points more useful than “build my app.”

The repository has expanded well past Markdown skills. It ships a browser daemon and Chrome interface, safety guards, edit locks, persistent project learning, optional GBrain integration, document generation, diagrams, model benchmarking, iPhone QA tooling, release helpers, hooks, generated host variants, and scripts for team installation. That breadth makes gstack powerful and also makes it software infrastructure that deserves normal dependency review.

Installation is fast, trust takes longer

The personal setup requires Git, Bun 1.0 or newer, a supported coding agent, and Node.js on Windows. It clones into the agent's skill directory and runs ./setup. The installer can target Claude Code, Codex, OpenCode, Cursor, Factory Droid, Slate, Kiro, Hermes, and GBrain. For Claude Code, the suggested prompt also adds a section to CLAUDE.md directing browsing through gstack and enumerating available skills.

Recommended team mode adds repository configuration and an hourly, silent auto-update check. Network failure does not stop work, and the check is throttled, but automatic upstream changes are still a supply-chain and reproducibility decision. A regulated or security-sensitive team should pin reviewed revisions, examine generated instructions, and control updates through its normal change process. “No version drift” is not automatically better than reviewed drift.

Capabilities have meaningful side effects. /ship can sync branches, run tests, push, and open a pull request. /land-and-deploy can merge and wait for production. Browser tools can import cookies from Chrome-family browsers, persist authenticated sessions, and open a tunnel for remote agents. Continuous checkpoints create commits, although pushing them is opt-in. Safety skills help, but they are additional agent instructions, not a substitute for OS isolation, scoped tokens, branch protection, and human approval.

Telemetry is opt-in and documented. The README says activation asks for consent, anonymous mode is available, prompts and code are not collected, and users can turn it off through configuration. Data goes to Supabase through validated functions. That disclosure is better than burying collection in setup, though teams should still verify current code and policy before deployment.

The current failures are not cosmetic

Issue 2519 reports at least 15.2 million Codex input tokens consumed during a review of a roughly 422-line diff. The parent spawned five review agents, invoked nested Codex review commands, and repeatedly sent very large overlapping contexts. Cached tokens may price differently, but the operational problem remains: an ostensibly bounded review produced runaway work within minutes. Cost controls and recursion guards should be tested before using gstack against large repositories.

Issue 2511 says two compatibility aliases install duplicate skill names under Claude Code. In the reported environment, Claude Code dropped the entire personal skill set, making every gstack command undiscoverable after restart. Issue 2502 finds the Codex and Copilot installer linking bin/ without the sibling lib/ directory required by 13 scripts. Errors are redirected or tolerated in preambles, so learning, decision, question-log, and telemetry writes can disappear silently. Failure visibility is as important as a fix here.

Windows has accumulated several precise portability reports. Issue 2195 shows a BSD-style stat -f probe succeeding with the wrong GNU meaning under Git Bash, leaving repository mode unknown. Other current reports cover hooks that never start because of Windows URL and path conversion, GBrain sync dispatched through cmd.exe instead of Bash, and setup behavior that copies files when Developer Mode does not allow symlinks. The README does warn Windows users to run through Git Bash or WSL and to reinstall copied skills after pulls, but parity is not yet boring.

Health is exceptional, stability is not

The repository was pushed on August 12, 2026, and same-day issue activity included fixes with focused regression tests. Its 934 open issues and pull requests must be read alongside 127,674 stars and intense maintainer and contributor traffic. This is not a neglected backlog. It is the pressure created when a young project adds hosts, operating systems, browser automation, deployment tools, and rapidly evolving model behavior at once.

GitHub's latest-release endpoint returned no release, even though the README refers to internal version numbers and current issues cite versions such as 1.61.0.0. A missing GitHub release does not indicate abandonment because current pushes and issue response are abundant. It does make reproducible acquisition less obvious than downloading a signed or checksummed release artifact.

Documentation is a major strength. The README explains requirements, team behavior, uninstall paths, privacy, platform caveats, and the purpose of each skill. Deep guides cover browser security and individual workflows. The sales voice is loud and productivity claims occupy too much opening space, but users can still discover what the tool changes.

Start smaller than the pitch

gstack is best for someone who already knows how to review an agent's work and wants a repeatable process. Begin with product discovery, planning, or report-only QA. Add mutation, shipping, cookie import, tunnels, and automatic updates only after observing each boundary. Set token budgets and watch subprocesses when using Codex review.

Do not confuse a collection of forceful prompts with an autonomous engineering organization. The quality ceiling remains the underlying model, repository context, test environment, and human judgment. gstack improves the process around those inputs, sometimes substantially, but its present integration bugs show that the process itself can fail. Use it as a sharp, evolving toolkit, not an unquestioned control plane.

Alternatives

ProjectWhat it isPick it when
SuperpowersA structured software-development workflow built from composable agent skills.Pick this instead when you want disciplined engineering skills with a smaller product-management and browser surface.
Anthropic SkillsReference and production skills published for Claude's skill system.Pick this instead when you want narrower examples and building blocks rather than one full operating system.
Andrej Karpathy SkillsA compact set of coding rules derived from Karpathy's stated agent practices.Pick this instead when concise constraints are preferable to an end-to-end automated sprint.

What people are saying

  1. [github-trending] garrytan/gstack

Sources

  1. gstack README
  2. Runaway Codex token report
  3. Duplicate Claude Code skill names report
  4. Missing Codex runtime library report
  5. Windows Git Bash repository mode report