Claude Forge 🤖 Agent Open source

What it does

Claude Forge is a large, opinionated setup pack for Claude Code that the author compares to oh-my-zsh. One install adds specialist subagents, slash commands, skills, hooks, rule files and a few MCP connections, all wired into a plan, test, review, verify and ship pipeline. Its main idea since version 4 is an adversarial verification loop: a second agent that did not write a change and never sees the first agent's reasoning checks it, and the loop repeats until that checker approves. The project is actively developed, with frequent releases and translated READMEs.

What is inside

• Agents: planner, architect, code, security and database reviewers, a TDD guide, build-error fixer, end-to-end test runner, dead-code cleaner and several debugging and verification roles
• Commands: workflows such as /plan, /tdd, /code-review, /verify-loop and /commit-push-pr, plus worktree, documentation and refactoring helpers
• Skills: reusable procedures including review-loop, loop-forge for turning repeated tasks into commands, and a context-budget checker
• Hooks: checks that block leaked secrets, destructive database commands and unsafe remote scripts, plus auto-resume after API errors, a loop guard and an end-of-session time report
• MCP: Playwright, Context7, a web page reader and Chrome DevTools

Works with

Built for Claude Code. The session time report hook can also read Codex transcripts if you wire it into Codex's hooks file by hand.

How to install

The plugin route gives commands and most skills only:

/plugin marketplace add sangrokjung/claude-forge
/plugin install claude-forge

The full install adds agents, hooks, rules and MCP servers:

git clone --recurse-submodules https://github.com/sangrokjung/claude-forge.git
cd claude-forge
./install.sh

Maintenance and safety

The repo is MIT-licensed. The full installer writes into ~/.claude, so back up an existing configuration first; an --upgrade flag is provided for older installs. The README also offers a pipe-to-shell one-liner from the project's own GitHub repo; cloning and reading the script first is the safer route.

Who should use it

Claude Code users who want a complete, pre-wired workflow with strong review and safety defaults and are comfortable adopting someone else's conventions. Anyone with a carefully tuned setup may prefer to cherry-pick parts.