gstack
TL;DR
gstack is an MIT-licensed set of Markdown slash-command skills for Claude Code, written by Garry Tan, that assigns roles such as product reviewer, engineering manager, QA lead and release engineer along a plan, build, review, test and ship sequence. It also ships a browser tool, and setup targets Codex, OpenCode and other hosts.
Key facts
| Type | Orchestrator |
|---|---|
| Languages / SDKs | TypeScript, Markdown, Shell |
| License | MIT |
| Pricing model | Open source, free |
| Orchestration pattern | Pipeline |
| GitHub stars | 134,550 (as of 2026-09-30) |
| GitHub forks | 20,031 |
| Last push | 2026-09-30 |
| Latest release | No release published |
| Repository | garrytan/gstack |
| Documentation | github.com |
| Last verified | 2026-09-30 |
Key features
- A sprint of role-named slash commands (
/office-hours,/plan-ceo-review,/plan-eng-review,/review,/qa,/ship,/document-release) in which each step reads the previous step's output. (source) /autoplanruns the CEO, design, DX and engineering plan reviews in sequence and picks which apply. (source)- Per-host setup (
./setup --host <name>) for Codex CLI, OpenCode, Cursor, Factory Droid, Kiro, Slate, OpenClaw, Hermes and GBrain, with an instruction-only digest for other agents. (source) - A browser layer: skills drive the user's Aside browser first, falling back to a bundled headless Chromium daemon, plus a sidebar agent that runs a child Claude instance. (source)
/pair-agentlets a second agent (OpenClaw, Hermes or Codex) join the same GStack Browser window in its own tab using a one-time setup key. (source)/codexand/claude-codesend code reviews to the other vendor's CLI for a second opinion, with session continuity. (source)- OpenClaw integration: OpenClaw spawns Claude Code sessions over ACP and tells them to load gstack skills; four methodology skills also install natively through ClawHub. (source)
- Safety commands
/careful,/freezeand/guard, and an opt-in verification stop hook that blocks a turn from ending until a declared verify command passes. (source)
Architecture and orchestration pattern
Pattern: Pipeline
gstack is a directory of Markdown skills plus TypeScript helpers, installed under ~/.claude/skills/gstack for Claude Code or generated into each other host's skills directory by ./setup. The skills are the orchestration: a sprint sequence in which /office-hours writes a design doc that /plan-ceo-review reads, /plan-eng-review writes a test plan that /qa picks up, and /ship runs the release steps. The host agent executes each skill; gstack has no scheduler.
Multiple agents appear in three places. Running several sprints in parallel is left to an external tool, Conductor, which the README says runs Claude Code sessions in isolated workspaces. /pair-agent and the OpenClaw integration coordinate separate agents through a shared browser or through OpenClaw spawning Claude Code sessions. Outside reviews call the other vendor's CLI. The sidebar agent is a child Claude instance in an isolated session.
State is local. Skills hand results to each other through design and test-plan files, a browser daemon keeps tabs and cookies between calls, and an optional GBrain knowledge base gives cross-session memory, registered with Claude Code as an MCP server when the user chooses it.
Human in the loop
Several skills are interactive: the design review asks one question per design choice, and the README says risky documentation changes return for approval. /careful warns before destructive commands, /freeze restricts edits to one directory, and an opt-in verify hook holds a turn open until a project's declared command passes, yielding after three blocked attempts. For the browser, a handoff command opens a visible Chrome at the same page when the agent hits a CAPTCHA or login, and resumes after the user finishes. Telemetry is opt-in and every off-machine send is logged to a local receipt file.
Harnesses it can drive
- Claude Code (evidence)
- Codex (evidence)
- OpenCode (evidence)
- OpenClaw (evidence)
- Hermes (evidence)
Protocols
| Protocol | Support | Note |
|---|---|---|
| MCP | Partial evidence | Client side only and optional: /setup-gbrain can register the separate GBrain server with Claude Code (claude mcp add gbrain -- gbrain serve) or point at a remote gbrain MCP URL; gstack ships no MCP server of its own and its CLAUDE.md snippet steers browsing to /browse instead of MCP browser tools. |
| A2A | Unknown | No A2A / Agent2Agent mention in the README or docs/OPENCLAW.md; repo code search hits were fixtures, lockfile and UI code, not protocol support. OpenClaw integration uses ACP, a different protocol. |
| AG-UI | Unknown | No AG-UI mention in the README; repo code search for ag-ui returned nothing; not checked against a listing in the AG-UI README. |
Best for
- Claude Code users, especially founders and tech leads, who want a fixed think, plan, build, review, QA and ship sequence run through slash commands. (shortlist)
- Adding browser-based QA and design review to a coding agent, using the user's own logged-in browser or a bundled Chromium. (shortlist)
- Running several Claude Code sessions in parallel under an external workspace manager, with the same process in each session. (shortlist)
- OpenClaw setups that need a planning and review discipline injected into the Claude Code sessions they spawn.
Not for
- Building agents in code; gstack is skills, CLIs and a browser daemon for existing coding agents.
- Automated parallel-agent scheduling or worktree management; the README defers that to Conductor.
- Machines without Bun or a build toolchain; setup requires Bun 1.0 or newer, and /cso needs extra Bun build flags and a native compiler.
Quickstart
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup # Claude Code: install
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack \
&& cd ~/.claude/skills/gstack && ./setup
# Other agents: clone anywhere, let setup detect hosts or name one
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/gstack
cd ~/gstack && ./setup --host codex
# Then, inside the agent
/office-hours # describe what you are building
/plan-eng-review # lock architecture and tests
/review # review the branch
/qa # test against a staging URL
/ship # release steps
Common pitfalls
- Requires Claude Code, Git and Bun 1.0 or newer (Node.js on Windows only).
/csoneeds a Bun release with four--no-compile-autoload-*build flags and a native toolchain; without them setup skips the CSO helpers and/csoreportsnot assessed.- The README's install prompt asks Claude to add a gstack section to CLAUDE.md that forbids the claude-in-chrome MCP tools in favour of /browse.
- Open issues report a Playwright Chromium download failure aborting setup, a generated Codex ship skill over a 160 KB token ceiling, and a broken Cursor host setup.
/claudewas renamed to/claude-code; re-run./setup --host <name>to migrate managed installs.- Under Conductor, API keys are stripped from workspace environments, so set the GSTACK_-prefixed key variables described in the README.
Pros
- Each sprint step writes a document the next step reads, so the process is inspectable in files rather than held in one prompt. (source)
- Per-host setup with documented directories for Codex, OpenCode, Cursor and others, plus a 2 KB instruction-only digest for other agents. (source)
- Telemetry is off by default, and every off-machine send writes a hash-chained local receipt that can be audited with
gstack-egress. (source) - Safety and verification controls (
/careful,/freeze,/guard, verify gate) are documented as opt-in commands. (source) - MIT licensed with no paid tier described in the README. (source)
Cons
- No GitHub releases are published, so installs track the default branch and updates come from the auto-update check. (source)
- Setup is heavy: Bun, a browser build, and for /cso extra Bun flags and a C toolchain; an open issue reports a Chromium download failure aborting setup before skills land. (source)
- Support outside Claude Code has open defects in the tracker, for example a generated Codex ship skill above a 160 KB token ceiling. (source)
- Parallel sprints depend on Conductor, a separate tool; gstack itself does not manage workspaces or worktrees. (source)
- An open issue asks to reduce static skill context, because always-on skill frontmatter costs tokens every session. (source)
Alternatives
FAQ
Does gstack support MCP?
Only partly and optionally. /setup-gbrain can register the separate GBrain server with Claude Code as an MCP server. gstack ships no MCP server, and its install prompt tells Claude to use /browse instead of the claude-in-chrome MCP tools.
Does gstack run multiple agents?
Not on its own. The skills are roles followed by the host agent in sequence. Parallel sessions come from Conductor, which the README recommends, while /pair-agent and the OpenClaw integration connect separate agents through a shared browser or through OpenClaw spawning Claude Code.
Which agents does it work with?
Claude Code first. The README lists setup hosts for Codex CLI, OpenCode, Cursor, Factory Droid, Kiro, Slate, OpenClaw, Hermes and GBrain, and an instruction-only digest for others.
Is gstack free?
Yes. The repository is MIT licensed and the README says it is free. Telemetry is opt-in, and the optional Aside browser and any model API usage are separate.
How is it different from Superpowers?
Both are skill packs with a fixed workflow. gstack centres on named roles and slash commands plus a browser layer for QA, while Superpowers centres on a plan-and-subagent workflow that its skills trigger automatically.
Sources
- gstack README
- gstack ARCHITECTURE.md
- gstack x OpenClaw integration doc
- gstack releases page (empty)
- Issue 1901: Playwright Chromium download failure aborts setup
- Issue 2777: Codex ship skill above 160 KB token ceiling
- Issue 2616: reduce static skill context
- gstack README, install section
- garrytan/gstack GitHub repository