gptaku_plugins: 18 Claude Code plugins sold as answers to specific walls

A marketplace directory with a .claude-plugin directory at its root

The project describes itself as a Claude Code plugin marketplace, and the repository layout supports that reading rather than the more common one where a marketplace is a thin index pointing at separate repositories. There is a plugins/ directory, a .claude-plugin/ directory at the root, and the whole thing installs from a single git URL. The primary language recorded for the repository is Python, and the license is MIT.

That single-repository arrangement is the main structural decision here. A reader can clone one URL and see every plugin in one tree, which makes it easy to audit before installing and easy to read as a catalogue. It also means the plugin count in the README is the plugin count in the repository, with no separate repositories to audit individually.

The tree carries more than plugins. There are scripts/, shared/, tools/ and docs/ directories, an AGENTS.md and a CLAUDE.md at the root, a paseo.json, and a file called sangse-HANDOFF.md. Five README files cover English, Korean, Chinese, Japanese and Spanish, which is a detail worth noticing because it says something about who the catalogue is written for.

The last push was on 2026-09-08 and the repository is not archived. There are no GitHub releases attached to it, so the version history of the marketplace as a whole is not something you can read off the releases page.

Three commands, and an explicit order for which plugin goes first

Installation happens inside Claude Code rather than through a package manager. The README gives three steps, and the middle one carries an instruction that is easy to skip: install insane-search first, because the other plugins assume you have a working Claude Code session.

# 1) Add the marketplace once
/plugin marketplace add https://github.com/fivetaku/gptaku_plugins.git

# 2) Start with the lowest-friction flagship
/plugin install insane-search@gptaku-plugins

# 3) Apply without restarting the whole app
/reload-plugins

The third step is the part that separates this from a restart-based workflow. Reloading plugins means you do not have to bounce the whole application to pick up a change, which matters more than it sounds if you are iterating on a plugin during a long session.

Each individual plugin then installs with the same shape of command, naming the marketplace you just added:

/plugin install insane-design@gptaku-plugins

The @gptaku-plugins suffix is the marketplace identifier, so it stays the same no matter which of the eighteen you pick. Nothing in the README requires you to install all of them, and the phrasing throughout is pitched at picking one.

The insane-* four are pitched at four different walls

The README groups the catalogue into a flagship four and then a longer routing table. The flagship four are named for what they get you past.

insane-search is aimed at pages Claude Code cannot read. The README describes it escalating through public API readers, syndication gateways, TLS impersonation and finally a real headless browser, trying each in turn, and claims no API keys are needed. The suggested prompt is a Reddit summarisation task.

insane-design reads the real CSS of a site and extracts what matters from it: colour, type, spacing, radius, shadow and fonts. The output is described as a clean token set, a design.md spec, and a click-to-copy HTML report. The example prompt is extracting the design system from a named URL.

insane-review is the unusual one, because it depends on a consumer web session rather than an API. The README states that GPT-5.5 Pro has no API, then explains that this plugin packs relevant code with repomix, runs it through a logged-in ChatGPT Pro web session, and returns a line-by-line review. It requires an active logged-in Plus or Pro session and no API keys.

insane-research is described as a seven-phase pipeline: scope the question, spawn specialised research agents, grade and cross-check every source, then compile a structured Markdown report with citations.

What unites the four is that each one names a specific blockage. That framing is the catalogue's real editorial contribution, and it is more useful than a category list would be.

The trust table is where the real cost of each plugin sits

Most of the README's most useful information is in a table under Trust and setup, which sorts the eighteen plugins by what they need from you rather than by what they do. The claim repeated in the install section is that most plugins need only a running Claude Code session and no additional API key.

The exceptions are then grouped into three. Google Workspace OAuth covers nopal, which needs a Google account and a one-time login validation through the gws CLI. A logged-in browser session covers insane-review, which needs an active logged-in ChatGPT Plus or Pro web session and explicitly no API keys. An external CLI covers pumasi, which needs the Codex CLI installed.

This is the part of the README that deserves a careful read before installing anything, because two of those categories hand a plugin access to an authenticated session. insane-review routes code through a logged-in consumer chat session, which is a different trust decision from calling a documented API endpoint with a scoped key. It may well be a reasonable trade for a review pass, but it is a trade, and the README is upfront about it rather than burying the requirement in a footnote.

The routing table above it covers the remaining plugins: show-me-the-prd for turning a vague idea into a product spec, goaljaby for a PRD that needs a verified /goal execution contract, pumasi or kkirikkiri for splitting a large build across workers, sangse for a selling detail page with an image cut sheet, dd for a screenshot, long log or copied reference, docs-guide for library and API questions that need official docs, nopal for Google Workspace services, git-teacher, vibe-sunsang for improving how you work with Claude Code, and skillers-suda for creating your own plugin, skill or agent.

Beyond the plugin list, the repo carries agent instructions and a handoff file

Some of the tree entries have nothing to do with the eighteen plugins, and they hint at how the project is built. AGENTS.md and CLAUDE.md at the root are agent instruction files, meaning the repository is developed with coding agents in the loop rather than only shipping them. sangse-HANDOFF.md suggests a handoff document kept between sessions for one plugin. paseo.json sits at the root with no explanation in the README.

None of that is explained, which is a fair criticism. The README documents installation, plugin behaviour and credential requirements in reasonable detail, then stops. There is no architecture note, no contribution guide, and no compatibility statement naming which versions of Claude Code the plugins were written against.

For a catalogue of this size that is a workable arrangement, because the README is the product. But it does mean that evaluating individual plugins means reading whatever lives under plugins/ for that one entry, and there is no per-plugin versioning to pin against. With no GitHub releases either, a plugin update arrives whenever the branch moves.

Where this fits against writing the plugin yourself

The honest comparison is not with another marketplace, it is with the plugin authoring tools the same ecosystem offers. The catalogue lists skillers-suda as a plugin for creating your own plugin, skill or agent, which is the sharpest framing of the choice: install a plugin for the problem you have, or install a plugin that teaches you to build one.

The case for the catalogue is coverage of awkward edges. Writing a plugin that escalates through syndication gateways and TLS impersonation and a headless browser is not a weekend project, and nobody would recommend rebuilding it for a single task. The same goes for a design token extractor that reads live CSS. Those are exactly the plugins where the shared implementation is worth more than the configuration.

The case against is dependency weight for a solved problem. Eighteen plugins that all load into one Claude Code session is eighteen sets of instructions in context, and the README offers no guidance on how they interact or whether they conflict. It also does not claim to be more than a collection of independent tools.

One structural risk deserves naming. The four flagship plugins lean on external surfaces: a WAF-bypassing fetch chain, a scraped design system, a consumer web session, and a research pipeline. All four can break without this repository changing, because what they depend on is not under its control. That is not a criticism of the work, it is the shape of the approach.

Editorial conclusion

This repository is a directory of answers rather than a framework, and reading it that way is the fastest way to decide about it. If your problem is a blocked page, a design system you cannot name, or a build that keeps stalling, the install path is three commands and the README names the plugin. If you need release notes, a changelog, or a stated compatibility range for a particular Claude Code version, none of that is here: the repository lists no GitHub releases, and the last push was on 2026-09-08. Start by adding the marketplace and installing insane-search, then read the Trust and setup table closely, because that table, not the plugin count, is where the real cost of each plugin shows up.

Frequently asked questions

Official sources

1. fivetaku/gptaku_plugins on GitHub
2. Issues
3. License: MIT
4. README