Developer releases MIT-licensed kit enforcing permission-first workflow for Claude Code with 10 agents and 28 skills.
Developer Sabahattin K built Full Stack HQ, a permission-first configuration kit for Claude Code and Google Antigravity IDE. The MIT-licensed kit adds 10 specialist agents and 28 skills to prevent unauthorized file modifications.
Key facts
- 10 specialist agents included.
- 28 skill modules for frameworks and tools.
- Only 4 approval keywords accepted.
- MIT-licensed, free to use.
- Install via single curl command.
The kit, published on GitHub under an MIT license, addresses a common complaint among Claude Code users: the agent executes actions without explicit approval. According to the developer's blog post, the agent 'deleted files I didn't want deleted' and 'refactored things I didn't ask it to refactor.'
Full Stack HQ enforces a strict workflow: the agent first presents a phased plan, waits for specific approval keywords (PLAN APPROVED, IMPLEMENTATION APPROVED, PROCEED, DO IT), then executes one phase at a time. The developer claims this 'eliminates 80% of unwanted surprises.'
What's inside
The kit installs into ~/.claude/ and ~/.gemini/ directories, adding:
- 10 specialist agents (frontend, backend, database, architect, code-reviewer, test-engineer, security-auditor, performance-optimizer, devops-engineer, documentation-writer)
- 28 skill modules covering frameworks like Next.js, NestJS, Prisma, and testing tools like Vitest and Playwright
- 10 slash-command workflows (
/plan,/brainstorm,/debug,/create,/enhance,/test,/orchestrate)
Installation is a single command: curl -fsSL https://raw.githubusercontent.com/sabahattink/antigravity-fullstack-hq/main/install.sh | bash with flags for --only-claude or --only-antigravity.
The unique take
The core insight is that Claude Code, like most agentic coding tools, lacks built-in permission granularity. Anthropic's own documentation for Claude Code emphasizes speed and direct file access — the product competes with Cursor and GitHub Copilot on raw throughput. Full Stack HQ reverses that philosophy: slow down, ask first, execute only on explicit approval. This mirrors a broader tension in the AI coding agent space between velocity and safety. The developer's framing — 'senior engineers, not interns with root access' — captures the trade-off succinctly.
Limitations
The kit is a CLAUDE.md file plus supporting configs, not a plugin or API-level enforcement. A determined user can override the rules. The developer notes the project is open to PRs, suggesting it's early-stage: the HN post has 4 points and 0 comments as of publication.
What to watch
Watch for Anthropic's official response — if Claude Code adds native permission controls in a future update, it would validate the approach and potentially make third-party kits obsolete. Also watch for adoption metrics: the GitHub star count and HN engagement will indicate whether the broader Claude Code community sees this as a must-have or a niche workaround.
[Updated 15 May via devto_claudecode]
The developer also published a Windows PowerShell installation command (irm ... | iex), expanding access beyond the Mac/Linux-only curl command previously covered [per dev.to post]. Additionally, the post lists 10 slash-command workflows (not 9 as earlier reported), adding /brainstorm and /debug to the suite.
Originally published on gentic.news
Top comments (0)