badi

BADi: A Workflow Management System for Claude Code

Language / Dil: English · Turkce

Claude Code does what you tell it — until it doesn't. badi is the agentic safety layer for Claude Code: 14 deterministic hooks that run as real code on every tool call and block the irreversible actions — a force-push to main, an rm -rf /, a .env piped to curl — before they execute. Not a prompt the model can reason around — code that intercepts the call. And it's fail-safe: if a hook ever errors it exits cleanly, so safety never wedges your session.

npx @fatihkan/badi init && badi doctor   # ~2 minutes

Beyond the guardrails, badi is a full Claude Code workflow layer — a daily ritual, code review, a virtual eng team, content & SEO — as 30 subagents, 86 commands, and 63 opt-in skills you enable only when you need them. Works with Cursor, Gemini CLI, and Windsurf too. Built for Claude Opus 4.7 / Sonnet 4.6 · 1321 passing tests · MIT.

Demo

Render the demo: brew install vhs && vhs assets/demo.tape produces assets/demo.gif deterministically. The tape is text-checked into the repo so the GIF is reproducible.

Why Badi?

You could hand-roll your own slash commands and agents — many people do. Badi is the maintained, tested alternative, and it starts with guarantees, not suggestions:

• Enforced, not suggested — guardrails run as code on every tool call: branch-guard blocks direct commits and force-pushes to protected branches, a bash guard hard-blocks destructive commands (rm -rf /, curl | bash, .env exfiltration), and a completeness gate blocks file writes that contain live secrets. A CLAUDE.md rule is advice; a hook is enforcement.
• Fail-safe by design — if a hook ever errors it exits cleanly and never wedges your session, so safety doesn't become friction.
• Curated, not assembled — 30 subagents and 86 commands that already work together (a virtual eng team, code review, content, mobile/SEO), instead of writing and maintaining each one yourself.
• One source, five harnesses — author once in .claude/; compile to Claude Code, Cursor, Gemini CLI, Windsurf, and AGENTS.md.
• Opt-in by default — skills load zero tokens until a prompt needs them, and profiles hide the commands you don't use.
• Tested — 1321 passing tests, with releases gated on a docs/security sync check.

Install Options (v1.30.1+)

Channel	Command	Status
npm (primary)	npm i -g @fatihkan/badi	✅ Live
Claude Code marketplace	/plugin install fatihkan/badi	✅ Live (manifest in repo)
Homebrew (macOS / Linux)	planned — tap repo not yet published; use npm	⏳ Planned
Scoop (Windows)	planned — bucket repo not yet published; use npm	⏳ Planned

Source of truth: npm. Other channels mirror the same @fatihkan/badi-<version>.tgz. Sha256/hashes are populated by the release workflow (.github/workflows/dist-publish.yml). Homebrew and Scoop channels are skeletoned in dist/ but the tap/bucket mirror repos (fatihkan/homebrew-badi, fatihkan/scoop-bucket) are not yet created — until then, use npm or the Claude Code marketplace.

One-Command Install

Pick your path: Quick — install as a Claude Code plugin (no Node.js). Full — install the npm CLI for everything (hooks, the multi-harness compiler, analytics, and the badi toolchain). Both paths ship the 30 agents, 86 commands, and 63 skills.

As a Claude Code plugin (no Node.js required for install):

# Inside Claude Code
/plugin marketplace add fatihkan/badi
/plugin install badi@badi-marketplace

As an npm CLI (full feature set: 30 agents · 86 commands with profile management (v1.26+) · 14 hooks · 63 opt-in skill categories with auto-router):