Claude Forge: an equipment pack for Claude Code, with an adversarial review loop

What Claude Forge actually adds to a bare Claude Code session

Claude Code is capable out of the box and arrives with no procedures attached. Claude Forge is the equipment pack layered on top: 16 specialist agents it can delegate to, 35 one-word slash commands, 33 skills, 22 hooks, 14 rule files and 4 external tool connections. The README frames this with an oh-my-zsh analogy, which is accurate in one respect and misleading in another. Accurate because, like oh-my-zsh, it changes nothing about what the underlying tool fundamentally does. Misleading because oh-my-zsh configures a shell, while Claude Forge configures an agent that edits your files and can run commands. The blast radius is different.

The intended audience is someone already using Claude Code in a terminal who keeps re-explaining the same expectations: write tests, check for secrets, update the docs. The pitch is that plan, test, review, verify and ship become connected steps rather than reminders you issue by hand. The repository is Shell, MIT licensed, and the last push was on 2026-09-03.

The adversarial verification loop and why maker and checker are separated

The v4.0.0 release introduces an adversarial verification loop. Every behavioral change is checked by a second, independent agent that did not write the code and never sees the first agent's reasoning. The loop runs until that second agent issues APPROVE. This is the most interesting design decision in the project, because it addresses a real failure mode: an agent reviewing its own work tends to accept the reasoning it just produced. Denying the checker access to that reasoning is the mechanism, not a stylistic preference.

The README states that while building v4.0 itself, the loop caught three real defects in the maintainer's own pull requests, numbered #58 and #61, and points to docs/VERIFICATION-LOOP.md for a worked example. Treat that as the maintainer's claim about their own process rather than an independently measured defect rate. The same release ships a reliability pack: API-error auto-resume, session relay across /compact, and a doom-loop guard, documented in docs/RELIABILITY.md. Task-grade routing decides how much machinery a given task gets. A Korean prose quality pack is included as well, reflecting the maintainer's own language.

Installing Claude Forge and running a first slash command

There are two install routes and they do not deliver the same thing. Option A is the plugin route. Open a Claude Code session and run two commands:

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

The README is explicit that this gets you commands and most skills right away, while agents, hooks, rules and MCP connections need Option B. Later updates go through /plugin update claude-forge. If you have seen a failed to load error mentioning Hook load failed: expected record, received undefined, the README says that bug is fixed as of v4.0.0 and links issues #52 and #57.

Option B is the full install, described as one line in your terminal. The README text available here truncates that line mid-URL, so read INSTALL.md for the complete command rather than copying a partial one. The repository also ships install.ps1 for Windows alongside install.sh. The README claims a 5-minute install and states that updates are a git pull. Verify both against INSTALL.md for your platform before you commit to the full route.

Once installed, the workflow commands are the entry point. The README gives /plan, /tdd and /code-review as examples of one-word shortcuts that trigger full workflows. A first real use is to run /plan on a task you would otherwise describe in a paragraph, then follow the pipeline it proposes rather than jumping straight to edits. The beginner walkthrough linked from the top of the README is advertised as a 60-second setup, available in 13 languages on the project's documentation site.

The hook layer is where Claude Forge can bite you

The README describes a 6-layer hook system that blocks risky actions before they happen, and lists 22 hooks in the current count. Two of them deserve scrutiny. The first is the unattended API-error auto-resume: when a session hits an API error, the hook resumes it without a person present. That is convenient and it also means a session can continue working after you have walked away. The second is session relay across /compact, which carries context through a compaction boundary. Both are the kind of behavior you want to read about in docs/RELIABILITY.md before enabling, not discover from a diff.

The v4.2.0 release adds session-time-report as the 22nd hook. It measures each session as it closes and reports at your next start when the session ended, how long it actually ran, and how much got done. The README notes it is resume-aware, so a session picked up days later reports that sitting rather than a bogus 400-hour total. The v4.3.0 release extends the same hook to read Codex rollout logs as well as Claude Code transcripts, which requires adding two entries to ~/.codex/hooks.json. That is a notable scope expansion: the project now instruments a second agent harness, and the wiring for it lives in your Codex configuration rather than in Claude Forge's own settings.

If you are not comfortable with hooks that act on a session while you are absent, the plugin route is the safer starting point precisely because it leaves hooks out.

Context budget is the constraint the project itself now manages

The v4.1.0 release is titled harness-diet and adds a 33rd skill whose stated job is to measure your always-loaded context, meaning CLAUDE.md plus rules, and shrink it back under budget without losing governance. The fact that this exists is informative. A pack with 14 rule files and a large skill library consumes context on every turn whether or not those rules apply to the current task, and the maintainer evidently hit that ceiling.

This is the honest trade-off of the whole approach. You gain a pre-wired pipeline and lose headroom. The harness-diet skill is the mitigation, not a refutation. If you already run a tight CLAUDE.md, adding 14 rule files is a real cost you should measure rather than assume away. Run the harness-diet skill early, before you have customized anything, so you have a baseline to compare against after you edit the rules.

Where a lighter setup is the better answer

The closest alternative in spirit is assembling your own configuration: a CLAUDE.md with your conventions, a handful of subagent definitions in .claude/, and two or three hooks you wrote and understand. That approach has no install script, no update path, and no 22-hook surface. It also has no adversarial verification loop unless you build one, and building one is genuinely hard, which is the strongest argument for borrowing this project's version.

The difference in approach is ownership. Claude Forge makes decisions for you across agents, commands, skills, hooks, rules and MCP connections, and gives you a git pull to keep them current. A hand-rolled setup makes you decide each of those, and you will likely decide to have fewer of them. Neither is universally right. If your work is a single language in a single repository with conventions you can state in one page, a hand-rolled CLAUDE.md is probably enough and the extra agents are overhead. If you move between repositories, languages and review requirements, the pre-wired pipeline earns its context cost.

Licence, maintenance and what an upgrade costs you

Claude Forge is MIT licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are preserved. That is a permissive licence and it is not legal advice; if you redistribute a modified copy inside a product, read the LICENSE file and take your own counsel. A practical implication of MIT combined with a dotfiles-style install is that forking is cheap if the maintainer's direction stops matching yours.

On maintenance, the last push was on 2026-09-03 and the repository is not archived. The release cadence visible here is roughly monthly across mid-2026: v3.1.0 in June, v4.0.0 and v4.1.0 in August, then v4.2.0 and v4.3.0 in September. The upgrade cost is not the pull itself, it is the re-reading. MIGRATION.md and MIGRATION.ko.md exist precisely because releases change behavior, and the v4.0.0 notes describe a reliability pack and routing changes that alter how sessions behave rather than merely adding files. Budget time for MIGRATION.md on every major bump, and expect the hook count to keep growing: it went from 21 events in the v3.0.1 alignment to 22 hooks by v4.2.0.

Editorial conclusion

Adopt Claude Forge if you already run Claude Code daily and want a pre-wired plan, test, review and verify pipeline without writing your own agent definitions. Skip it if you want a minimal harness you fully understand, or if you cannot accept hooks that act on your session unattended. Before installing, read hooks/README.md and docs/RELIABILITY.md, and check whether the always-loaded context measured by the harness-diet skill fits your budget. The one thing to verify on your own machine is the full install path in INSTALL.md, because the README truncates the curl line and the plugin route deliberately ships only commands and most skills.

Frequently asked questions

Official sources

1. License: MIT
2. Project website
3. README
4. Releases
5. sangrokjung/claude-forge on GitHub