Claude Forge installs 22 hooks by piping a script into bash, and its counts disagree with its own description
oh-my-zsh for Claude Code — 16 agents, 35 commands, 32 skills, 21 safety hooks in one install. v4.0 adds an adversarial review loop: a second agent that never sees the first one's reasoning. MIT.
At a glance
- What is it?
- Claude Forge is a configuration kit for the Claude Code terminal assistant, sold with an oh-my-zsh analogy and delivered by one install line. The substance is a second agent that reviews work it did not write, a set of hooks that run silently in the background, and a hook that reports how long your sessions actually lasted. The housekeeping is sloppier: the project description still says 21 hooks and 32 skills, and the page advertises a version with no release attached.
- Who is it for?
- This kit suits someone who runs Claude Code all day and wants guardrails and a second opinion without wiring them by hand, and the adversarial loop is the part other kits should copy. Two cautions.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 73 days ago.
- What is it written in?
- Mainly Shell, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- sangrokjung/claude-forge881 ★ · +44
| Repository | 2026-09-11 | 2026-11-15 |
|---|---|---|
| sangrokjung/claude-forge | 837 | 881 |
44 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The recommended install pipes a remote script into your shell
The full install is one line, and it is a download executed directly:
curl -fsSL https://raw.githubusercontent.com/sangrokjung/claude-forge/main/install.sh | bashThe clone path is offered for people who would rather look first, and it carries a detail that matters: the clone command includes a flag for submodules, which tells you the repository has dependencies of its own that a plain clone leaves empty.
git clone --recurse-submodules https://github.com/sangrokjung/claude-forge.git
cd claude-forge
./install.sh # fresh install
# or
./install.sh --upgrade # safe migration from v2.1Windows users are told to run a PowerShell script as Administrator, so the kit expects to write into places a normal user process may not reach. What it writes is visible in the tree: a settings file, an MCP configuration, and directories for hooks and rules all sit at the repository root, waiting to be merged into a configuration you already have.
The plugin route gives you commands and half the skills, and it broke for a while
There is a lighter path through Claude Code's own plugin system, two commands in a session:
/plugin marketplace add sangrokjung/claude-forge
/plugin install claude-forgeWhat you get is stated without hedging: commands and most skills right away, while agents, hooks, rules, and the MCP connections require the full install. The comparison table makes it a grid of ticks, with skills marked partial and the other four categories marked unavailable. So a plugin user is running a subset, and the subset is the part that does not touch your files. That matters because of the next detail. The page carries a note that if the plugin install ever showed a failed to load message with the text Hook load failed, expected record, received undefined, that bug was fixed as of v4.0.0, with two issue numbers attached. Plugin users were getting a broken load and the fix came a full major version later.
The project description is a release behind the page, and the page is ahead of the releases
Two counting problems are visible without reading any code. The project's own description lists 16 agents, 35 commands, 32 skills, and 21 safety hooks. The page lists 35 commands, 33 saved procedures, 22 safety checks, 14 rule files, and 4 external tool connections. So the skill and hook counts in the summary a visitor sees first are each one behind the current state, which suggests the description was written at v4.0 and not updated for the two hooks and skills added since. The second problem is version numbering. The most recent tagged release is v4.1.0 from late August 2026, while the top of the page announces v4.3.0 as the latest, dated September 2026, and describes v4.2.0 as well. Neither of those two has a release attached, so anyone installing from a tag and anyone following the page are on different builds.
The adversarial loop refuses to approve until a second agent signs off
The headline feature of v4.0 is a review loop with one rule: the reviewer did not write the code and never sees the writer's reasoning. The page compresses the property into an inequality, maker is not checker until the checker issues an approval, which is a stronger requirement than asking an agent to self review, since self review is exactly the case where the reasoning that produced the defect also explains it away. The evidence offered is the loop applied to its own construction: while building v4.0 it caught three real defects in the maintainer's own pull requests, with two issue numbers given, and there is a worked example document in the repository. The same release ships a reliability pack and a debugging escalation chain, and adds task grade routing, which is how a request gets assigned a grade before an agent picks it up.
Twenty two hooks run silently, and one of them times your day
The hooks are described as safety checks that run silently in the background every time the assistant touches your code, and the inventory names three: an unattended API error auto resume, a doom loop guard, and an end of session time report. That last one is the newest, added as the twenty second hook, and its description is the most human thing on the page. It measures each session as it closes and tells you at your next start what time the session ended, how long it actually ran, and how much you got through. It is resume aware, so a session you picked back up days later reports the sitting time rather than a total of four hundred hours. The newest release then extended that same hook to a second assistant: it now detects whether a transcript belongs to Claude Code or to Codex, and the wiring for Codex is two entries you add to a file in your home directory.
One of its own skills measures the context your editor loads every session
harness diet arrived in v4.1 and is described as the thirty third skill. What it does is measure the always loaded context, named as the project instructions file plus the rule files, and shrink it back under budget without losing governance. That is a self referential addition to a kit that adds fourteen rule files of its own: the project ships a tool for noticing that its own kind of addition has grown too large to load cheaply. Earlier releases show the same attention to platform drift. v3.0.1 is described as alignment with a 2026 standard, listing hooks with 21 or more events, a second version of subagent frontmatter, a hybrid policy for skills and commands, and a four server minimum for the external tool layer. v3.1 added loop forge, which turns a repetitive task into a reusable slash command that guards itself.
The root holds two MCP manifests and a directory for a second assistant
The repository inventory is where the kit's ambitions show. At the top level sit a plugin manifest directory for the Claude Code path, a separate directory named for Codex, and both a required and an optional manifest for the external tool servers, alongside a third MCP configuration file. The page calls the external connections four, and an earlier release note calls four a minimum, which means the number is a floor rather than a total. There is also a knowledge directory, a libraries directory, a reference directory, a web directory, and two directories whose names the page never explains, both starting with the same prefix as the project. Documentation is triplicate for the readme, in English, Korean, and Chinese, while the migration guides come in only English and Korean, so a Chinese reader gets setup advice without upgrade advice.
Editorial conclusion
This kit suits someone who runs Claude Code all day and wants guardrails and a second opinion without wiring them by hand, and the adversarial loop is the part other kits should copy. Two cautions. The recommended install pipes a remote script straight into a shell and writes configuration into your home directory, so read install.sh first, or clone with the submodule flag the docs mention and run it yourself. And trust the counts on the page rather than in the project description, which still says 21 hooks and 32 skills.
Frequently asked questions
what is claude forge
It is an installable kit for Claude Code, described as an oh-my-zsh style equipment pack. One install adds 16 specialist agents, 35 slash commands, 33 skills, 22 background hooks, 14 rule files, and 4 external tool connections.
How do I install claude-forge?
As a plugin you get commands and most skills with two commands in a session. For everything, run the install script, either piped from the repository or after cloning with the submodules flag, where an upgrade flag handles migration from v2.1. Windows uses a PowerShell script as Administrator.
What does the adversarial verification loop in claude-forge do?
Every behavioral change is checked by a second independent agent that did not write the code and never sees the first agent's reasoning, and it must issue an approval before the change counts as done. While building v4.0 the loop caught three defects in the maintainer's own pull requests.
What is harness-diet in claude-forge?
It is the thirty third skill in the kit. It measures the context your editor loads on every session, the project instructions file plus the rule files, and shrinks it back under budget without dropping the governance rules.
Does claude-forge work with Codex as well as Claude Code?
Since the release described as v4.3.0, the session time report hook reads both, detecting whether a transcript is a Claude Code transcript or a Codex rollout log. Wiring for Codex is two entries added to the hooks file in your home directory.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/sangrokjung-claude-forge)Star history chart for your README
[](https://hysenlabs.com/projects/sangrokjung-claude-forge)Was this useful?