Claude Forge
A Claude Code toolkit that prewires specialist workflows, slash commands, and safety hooks.
- Source repo
- sangrokjung/claude-forge
- Stars
- ★ 875
- Last updated
- 2mo ago
- License
- MIT
- Primary language
- Shell
- FA score
- 55/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specificClaude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A team building a feature in Claude Code that wants an approved implementation plan from /plan before starting test-first work with /tdd.
- Not a fit if
- Teams standardized on ChatGPT, Codex, or the OpenAI API
- Teams that cannot send code and session data to a third-party service
- Source review
- 55/100 · Major gaps
What does this agent do, and when should you use it?
Claude Forge is a development-environment extension for Claude Code, delivered through either a plugin install or a fuller local installation. The full install provides 11 specialist agents, 34 slash commands, 26 skills, 15 built-in hooks, 10 rule files, and four default MCP connections. Its installer links repository configuration into ~/.claude/ on macOS and Linux so Claude Code can load the definitions, commands, rules, and hooks; Windows uses copied files instead. The documented primary workflow is /plan → /tdd → /code-review → /handoff-verify → /commit-push-pr. It fits Claude Code users who want a standardized plan-to-verification workflow, rather than teams seeking a standalone, provider-neutral agent API.
The installer connects agents/, commands/, skills/, rules/, hooks/, and MCP configuration to Claude Code. /plan creates an implementation plan and waits for approval; /tdd follows a failing-test, passing-test, refactor cycle; /code-review performs quality and security checks; and /handoff-verify runs build, test, and lint verification. Its agent definitions include planner, architect, security-reviewer, tdd-guide, database-reviewer, e2e-runner, and verify-agent. Built-in hooks run around Bash commands, database commands, file edits, and MCP use, including remote-command-guard.sh, db-guard.sh, and output-secret-filter.sh. The default MCP servers are playwright, context7, jina-reader, and chrome-devtools.
- A team building a feature in Claude Code that wants an approved implementation plan from /plan before starting test-first work with /tdd.
- A TypeScript maintainer who needs /build-fix or build-error-resolver to address build errors with minimal surrounding changes.
- An engineer preparing a release who wants /security-review and security-reviewer to check OWASP Top 10 issues, injection risks, and leaked secrets.
- A web developer who needs /e2e and e2e-runner to generate and run Playwright end-to-end tests.
- A team that wants a separate verification step to run build, tests, and lint through /handoff-verify before handoff.
- A Claude Code user repeating the same operation who wants /loop-forge to turn it into a reusable slash command with automatic verification and a hardstop.
How do you install or deploy this agent?
Prerequisite: Claude Code; the README badges Claude Code >=2.1.110. For the quick, partial plugin installation, run these inside a Claude Code session:
/plugin marketplace add sangrokjung/claude-forge
/plugin install claude-forgeThis provides commands and most skills; agents, hooks, rules, and MCP connections require the full install. On macOS/Linux, run:
curl -fsSL https://raw.githubusercontent.com/sangrokjung/claude-forge/main/install.sh | bashAlternatively, run git clone --recurse-submodules https://github.com/sangrokjung/claude-forge.git, enter the directory, and run ./install.sh. On Windows, run .\install.ps1 in PowerShell as Administrator. The supplied material does not document additional credentials.
How do you use this agent?
After a full install, start with /guide for an interactive tour. For a feature workflow, run /plan → /tdd → /code-review → /handoff-verify → /commit-push-pr; /auto login page is a documented example of starting an automated plan-to-PR flow. On macOS/Linux, run git pull in the claude-forge repository to update linked configuration; on Windows, rerun install.ps1 after pulling. To override settings without editing tracked files, copy setup/settings.local.template.json to ~/.claude/settings.local.json and edit it.
What are this agent's strengths and limitations?
- The full install bundles agents, commands, skills, rules, hooks, and four MCP connections as one prewired configuration rather than a single workflow.
- It documents a connected workflow spanning planning, test-first development, review, fresh-context verification, and commit/PR creation.
- Its hooks target unsafe remote pipelines, destructive SQL, secret leakage in output, and MCP usage, with additional opt-in lifecycle examples.
- macOS/Linux symlinks into ~/.claude/, allowing repository updates to take effect after git pull without reinstalling.
- Its core operation depends on Claude Code; no standalone adapter for ChatGPT, Codex, the OpenAI API, or the Claude API is documented.
- Plugin installation is intentionally incomplete: agents, hooks, rules, and MCP connections require the full installer.
- The default tool set includes Playwright, Context7, Jina Reader, and Chrome DevTools, so using those integrations requires network access and an MCP-capable environment.
- Windows uses file copies instead of symlinks, requiring install.ps1 to be rerun after each git pull.
How does this agent compare with similar options?
Against a basic .claude/ setup, Claude Forge supplies preconfigured specialist agents, commands, skills, hooks, and MCP connections. Against installing individual plugins, it positions /plan, /tdd, /code-review, /handoff-verify, and /commit-push-pr as a connected workflow.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Claude Forge This agent | 55 · Major gaps | Agent plugin / skillFree + model costs | ★ 875 | 2mo ago | Shell | Claude Code |
| Claude Code Slash Commands | 44 · Major gaps | Agent plugin / skillFree + model costs | ★ 2.6k | 1mo ago | — | Claude Code |
| Claude Code: Everything You Need to Know | 79 · Good | CLIFree + model costs | ★ 3.7k | today | TypeScript | Claude Code |
| OrchestKit AI Development Toolkit | 70 · Some gaps | Agent plugin / skillFree + model costs | ★ 300 | today | TypeScript | Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The README, hook documentation, and supplied tests show secret filtering, dangerous-command and destructive-SQL guards, rate limiting, restrictive log permissions, and symlink skipping. The packet does not establish the permissions or full data flows of every hook. Installation wires hooks, MCP servers, and shell configuration and includes automatic commit or resume behaviors, while confirmation and rollback guidance is limited; no broad dependency security review is shown. The manifest gives an author name and repository, but publisher identity remains unverified.
Regression cases cover auto-resume classification, the hook guard, and Korean punctuation checks, and CI validates manifests and formats. However, README inventory counts and version descriptions conflict: it describes 22 hooks for v4.3 while also describing 21 for v4.0. This weakens self-consistency. Some missing-runner and plugin-load failure cases are documented, but general diagnostics and dependency outage handling are only thinly evidenced.
The README provides workflows, command inventories, and scenario guidance for beginners, developers, and several task types. Skill descriptions include triggers and exclusions, and installation documentation distinguishes plugin and full installs. The sample covers only a few skills, and many capabilities depend on Claude Code, specific versions, Node, jq, Git, or experimental Agent Teams, limiting demonstrated environment fit.
The documentation is clearly organized and gives concrete install paths, platform notes, FAQs, and plugin limitations; the MIT license is complete. Manifests, CI version checks, and migration notes support versioning, but component counts and version descriptions are not fully consistent. SECURITY.md lists only 1.x as supported, and the packet does not independently corroborate its response commitments or establish maintenance responsibility, so full marks are not justified.
Commands, specialist agents, workflows, and safety hooks connect planning, implementation, review, and verification; severity-ranked findings and session reports are usable output forms. Claimed marginal gains are mostly project assertions rather than independently substantiated results in this packet. Automated hooks, MCP servers, and parallel agents also impose setup, runtime, and configuration costs.
The manifest, README, documentation, and several tests make some claims traceable, but tests cover only a small subset, and the install-validation document's reported run is not independent verification for this review. Major benefit and safety claims remain self-reported, with inventory and version inconsistencies; fact and inference are not always clearly separated.
- The full installer wires hooks, MCP servers, and shell settings into Claude's configuration. Review the scripts and permissions before installation, and check that automatic resume and commit-related behavior match your expectations.
- README counts for hooks, skills, and versions conflict. Confirm the actual files and manifests at the target revision.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/claude-forge)<a href="https://followagents.com/en/agents/claude-forge"><img src="https://followagents.com/badge/claude-forge.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v4.1.0 v4.1.0 — harness-diet: put your always-loaded context back under budget
Release notes excerpt
v4.1.0 — harness-diet: put your always-loaded context back under budget Your harness gains weight every week — this release ships the skill that takes it off without losing a single rule. What's new harness-diet (skill #33). A harness only grows: every incident note, every fix, every directive lands as another paragraph in an always-loaded rule. The methodology doc measured the asymmetry on a mature harness at roughly a 60:1 growth-to-reduction commit ratio, with natural growth around 70KB/month. This skill packages the diet that reverses it: …View on GitHub ↗ -
v4.0.0 v4.0.0 — adversarial verification loop, reliability pack, task-grade routing, Korean quality pack
Release notes excerpt
Nine capability sets ported from the maintainer's internal Claude Code harness. Selection was strict: demonstrated external demand (this repo's own contributor PRs and issues), the adversarial verification loop as flagship, and the Korean-community differentiator. Highlights Adversarial verification loop (flagship). Every behavioral change is checked by a fresh, independent reviewer that did not write the code — maker≠checker until the checker issues APPROVE. Not theoretical: building v4.0 itself, the loop caught three real defects in the maintainer's own PRs (#58, #61), and the relea …View on GitHub ↗ -
v3.1.0 v3.1.0 — loop-forge + beginner-friendly README
Release notes excerpt
v3.1.0 — loop-forge + beginner-friendly README ✨ New - loop-forge skill + command — turn a one-line repetitive task into a reusable, self-guarding slash command. It diagnoses the task into one of 5 loop shapes (Batch / Pipeline / Refine / Watch / Explore), interviews you for the blanks, and auto-injects two safety devices most people forget: an independent verifier (maker ≠ checker) and a hardstop (a budget/count/cooldown ceiling). Skills 24 → 25, commands 33 → 34. 📖 Docs …View on GitHub ↗
FAQ
Does the Claude Code plugin installation include every component?
What local locations does it modify or use?
What do the security hooks actually block?
Can the additional lifecycle hook examples be enabled selectively?
Thanks — we use this to improve the page.