- Category
- AI Agents
- Rank
- No. 35Tools index
- Listed in
- #7 Control a browser
- Pricing
- Open Source
- Type
- TOOL
- Builder
- @garrytan
- GitHub
- 132.1k stars
- Date
About
An end-to-end engineering skill system that builds a whole app end to end with Claude Code — from idea to a deployed production app. It takes a product all the way through planning, coding, automated review, testing, and deployment in eight commands, turning Claude Code into a virtual engineering team with 23 specialized roles (CEO, designer, engineer manager, QA lead, security officer) and structured slash-command workflows.
What it does
gstack packages engineering judgment as guided agent skills, then carries context between product discovery, architecture, implementation checks, browser testing, release work, and retrospectives. Its substantial runtime component drives Chromium for live interaction, screenshots, authenticated testing, performance checks, and remote agent access. Markdown supplies much of the methodology, while TypeScript and Bun support generation, setup, testing, browser control, and utilities.
Why it's ranked here
The appeal is unusually concrete: opinionated planning is paired with real browser automation, regression testing, security analysis, release handling, and persistent project learnings. The repository also describes extensive unit, end-to-end, and model evaluation surfaces. The trade is commitment. gstack imposes its own staged process, installs into agent-specific skill directories, and can add shared repository instructions plus automatic update checks.
What's good
The stages exchange useful artifacts instead of starting from blank prompts. Product discovery feeds planning, engineering review produces test material, and review findings reach release checks. Browser testing uses real Chromium and can verify fixes after making atomic commits. Host support is declarative, so adding an agent mostly means defining installation paths, metadata transformations, and tool rewrites rather than modifying each generator and setup component.
Tradeoffs
Installation requires Git and Bun, with Node.js additionally required on Windows. Team mode can require gstack for AI-assisted work and performs hourly throttled update checks, which may conflict with tightly pinned environments. Some integrations remain prompt bridges rather than shared runtimes. Cross-machine memory adds short network pauses around skill runs, and remote browser access depends on scoped tokens plus an ngrok tunnel when agents are on different machines.
How to use it well
It best suits technical founders, staff engineers, and small teams willing to standardize feature work around explicit discovery, reviewed plans, browser QA, security checks, and release verification. Start with one meaningful feature and stop after planning, review, and staging QA to judge fit. Keep existing CI and deployment controls authoritative. Treat gstack as engineering workflow and browser infrastructure, not as a replacement for messaging, calendars, or a general orchestration runtime.
Technical notes+
package.json defines an ESM Bun package with CLI binaries at browse/dist/browse and make-pdf/dist/pdf, Playwright and puppeteer-core dependencies, and Bun test commands spanning browse/test/, test/, and make-pdf/test/. docs/skills.md documents the skill surface. docs/ADDING_A_HOST.md describes typed host configs and generator-driven adaptation. docs/OPENCLAW.md explicitly frames that integration as prompt text with no daemon or JSON-RPC. docs/REMOTE_BROWSER_ACCESS.md specifies separate local and tunnel listeners, bearer-token scopes, tab ownership, an allowlisted command surface, and HTTP endpoints including POST /connect and POST /command. docs/gbrain-sync.md describes allowlisted Git-backed memory sync with credential scanning and custom merge drivers.
Observed
- License
- MIT
- Runtime and packaging
- ES module package requiring Bun 1.0 or newer; Node.js is additionally required on Windows
- Install surface
- Git clone plus a setup script, with global and team installation modes
- Interfaces
- Agent slash-command skills, browse and make-pdf CLIs, and a scoped HTTP browser API
- Browser stack
- Playwright-based Chromium automation, with puppeteer-core also declared as a dependency
- Host structure
- Declarative configurations are documented for Claude, Codex, Factory, Kiro, OpenCode, Slate, Cursor, and OpenClaw
- Test structure
- Bun test suites cover browser code, shared tests, PDF generation, agent routing, and model evaluations
Read from README.md, package.json, docs/skills.md, docs/OPENCLAW.md, docs/gbrain-sync.md, docs/ADDING_A_HOST.md, docs/domain-skills.md, docs/gbrain-sync-errors.md, docs/askuserquestion-cjk.md, docs/REMOTE_BROWSER_ACCESS.md, docs/askuserquestion-split.md, docs/gbrain-write-surfaces.md, docs/ON_THE_LOC_CONTROVERSY.md.
What it can do
Assign specialized AI roles to development tasks
Development task or workflow request → Task assignments to specific AI roles (CEO, designer, engineer manager, QA lead, security officer, etc.)
Execute product planning workflows
Product requirements or project specifications → Structured product plans and development roadmaps
Perform automated code review
Source code files or code changes → Code review feedback and improvement recommendations
Generate automated testing procedures
Application code and test requirements → Test plans and automated testing workflows
Execute deployment workflows
Application code and deployment configurations → Automated deployment processes and scripts
Process slash commands for development operations
Slash command with parameters → Executed development workflow or task completion
Coordinate multi-role development workflows
Complex development project requirements → Coordinated output from multiple AI specialist roles working together
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.