sangrokjung/claude-forge

Claude Forge – a plug‑in pack for Claude Code

What it is – Claude Forge is a ready‑made collection of agents, commands, skills, hooks, rules and tool integrations that turn the basic Claude Code terminal‑based coding assistant into a full‑featured development workstation. Think of it like oh‑my‑zsh for a shell: the core assistant stays the same, but the pack adds dozens of specialist “colleagues”, safety‑net checks and one‑line shortcuts.

Why it matters – Claude Code can generate code, but on its own you have to remember to run tests, do security reviews, keep docs in sync, etc. Claude Forge wires those best‑practice steps together automatically, adds a 6‑layer hook system that blocks risky actions (e.g., leaking secrets, running unsafe commands), and provides a verification loop where a second, independent agent reviews every change before it’s approved.

---

Key components (as listed in the README)

Component	Count	Role (plain language)
Agents	16	Specialist AI “colleagues” (planner, architect, security reviewer, test‑first guide, database reviewer, etc.) that are invoked automatically for the right task.
Commands	35	Slash‑style shortcuts you type in a Claude Code session (/plan, /tdd, /code‑review, /e2e, …) that launch whole workflows with a single keystroke.
Skills	33	Pre‑written playbooks that Claude follows without further prompting (e.g., loop‑forge, review‑loop, harness‑diet).
Hooks	22 (plus 9 optional examples)	Silent safety checks that run before/after every Claude action – they block secret leaks, dangerous DB commands, auto‑resume after API errors, report session time, etc.
Rules	14	Text files read at session start that define coding style, security policies, git workflow, when verification is mandatory, etc.
MCP connections	4	External tool bridges – Playwright browser automation, live docs lookup, Jina web‑page reader, Chrome DevTools integration.

---

Installation paths

Method	What you get
Option A – Claude Code plugin	Installs the 35 commands and most skills. Agents, hooks, rules and external‑tool connections are not installed.
Option B – Full install (recommended)	Runs a one‑liner `curl …

Both methods can be updated later with /plugin update claude-forge (Option A) or ./install.sh --upgrade (Option B).

---

How you use it

1. Start a Claude Code session in your terminal.
2. Type a slash command, e.g. /plan → Claude asks you to confirm the generated implementation plan.
3. Run /tdd → tests are written first, then code, with the tdd‑guide agent ensuring the RED‑GREEN‑REFactor cycle.
4. When you finish a change, invoke /code‑review → the code‑reviewer and security‑reviewer agents run, while the 22 hooks silently block any unsafe operations.
5. Finish with /handoff‑verify or /verify‑loop – a fresh‑context adversarial‑reviewer checks the change, returns APPROVE or REQUEST_CHANGES, and the session can be auto‑committed with /commit‑push‑pr.

All of the above is orchestrated by the built‑in team‑orchestrator agent, which can run several agents in parallel and route work to the most appropriate specialist.

---

Architecture snapshot (as described in the README)

• Core – Claude Code runtime that interprets slash commands and skills.
• Agent layer – 16 dedicated LLM agents, split into deep‑analysis (high‑capacity model) and fast‑execution (lighter model) groups.
• Hook system – 22 event‑driven safety checks that fire on 21 lifecycle moments (pre‑command, post‑command, error, session start/end, etc.).
• Rule engine – Static markdown files (CLAUDE.md + rule files) that are loaded at session start to steer the model’s behaviour.
• MCP (Multi‑Channel Plug‑in) connections – External services accessed via Playwright, Jina‑reader, Chrome DevTools, etc., exposing them as sub‑agents.
• Verification loop – A maker‑≠‑checker pattern where the adversarial‑reviewer (never sees the maker’s reasoning) must explicitly APPROVE a change before it is considered done.

---

Who might find this useful?

• Developers who already use Claude Code and want a production‑grade workflow without hand‑crafting their own safety nets.
• Teams that need reproducible, auditable AI‑generated code (security, compliance, SOC2/ISO27001, GDPR, HIPAA contexts).
• Solo engineers who want a “one‑click” pipeline from planning to PR, with built‑in testing, security review and documentation sync.

---

Quick start checklist

1. Open Claude Code.
2. Run the two plugin commands (Option A) or execute the install.sh script (Option B).
3. Try a demo command: /plan → confirm → /tdd → watch the generated tests and code.
4. When satisfied, run /commit‑push‑pr to create a PR automatically.
5. Keep the pack up‑to‑date with ./install.sh --upgrade or /plugin update claude‑forge.

---

License & community

• MIT License – free for personal or commercial use.
• The repo shows active maintenance (latest v4.3.0 released September 2026) and a growing contributor base (stars, forks, recent commits).
• Documentation includes multilingual setup guides (13 languages) and a detailed Verification Loop write‑up.

---

Bottom line: Claude Forge is a comprehensive, plug‑and‑play enhancement for the Claude Code AI coding assistant, delivering specialist agents, safety hooks, ready‑made commands and tool integrations so you can treat Claude as a fully‑equipped development teammate rather than a bare‑bones code generator.