plugins-for-claude-natives is a marketplace of Claude Code extensions for power users

Two slash commands and a marketplace manifest

plugins-for-claude-natives is a collection of Claude Code plugins for power users who want to extend Claude Code's capabilities beyond the defaults. The installation path is short enough to fit in two lines.

# Add this marketplace to Claude Code
/plugin marketplace add team-attention/plugins-for-claude-natives

# Install any plugin
/plugin install <plugin-name>

The first command registers the repository as a marketplace, and the second installs an individual plugin from it. The repository layout supports that model: there is a .claude-plugin/ directory at the root, which is where the marketplace manifest lives, and a plugins/ directory holding the individual plugin folders, one per plugin, with the README linking into each of them.

There is also an assets/ directory with the demo images the README shows for some plugins, and two versions of the README, the English one and README.ko.md. The Korean version is a hint about the project's centre of gravity, which is visible in the plugin list: several plugins are built around Korean users, from a KakaoTalk integration on macOS to a Korean podcast generator and a Korean translation step inside the YouTube digest.

The project is MIT licensed with a LICENSE file at the root, the default branch is main, and the repository publishes no GitHub releases. The last push was on 2026-04-20, so the tree is a moving target rather than a tagged distribution, which matters for a plugin set that runs inside somebody's coding session.

agent-council asks several models the same question

The agent-council plugin is the one with the clearest idea, and the README describes it as summoning multiple AI models to debate your question and reach a consensus.

When you are facing a tough decision or want diverse perspectives, the plugin queries multiple AI agents, specifically the Gemini CLI, GPT and Codex, in parallel and synthesizes their opinions into a single, balanced answer. The three steps it lists are: your question is sent to multiple AI agents simultaneously, each agent provides its perspective, and Claude synthesizes the responses into a consensus view with noted disagreements.

That last detail is the useful one. The output is not a vote and not a concatenation. It is a synthesis that keeps the disagreements visible, which is the part that makes multi-model consultation worth the tokens: the interesting content in a three model comparison is usually where they disagree.

The plugin is triggered by natural phrases rather than a command, with trigger phrases including "summon the council", "ask other AIs" and "what do other models think?".

# Example
User: "summon the council - should I use TypeScript or JavaScript for my new project?"

The practical requirement is visible in that list of agents. Gemini CLI, GPT and Codex are not bundled with the plugin, so each has to be installed and authenticated on your machine, and the plugin is a coordinator over CLIs you already run rather than an API client of its own.

clarify interviews you, doubt makes Claude argue with itself

Two plugins attack the same problem from opposite ends.

clarify transforms vague requirements into precise specifications through iterative questioning. Before you write code based on ambiguous instructions, the plugin conducts a structured interview to extract what you actually need. The process has four steps: capture, which records the original requirement verbatim, question, which asks targeted multiple choice questions to resolve ambiguities, compare, which presents the before and after showing the transformation, and save, which optionally writes the clarified spec to a file.

The example the README gives is a good illustration of the point. A requirement of "Add a login feature" becomes a spec with a goal, a scope of login, logout, registration and password reset, and constraints including a 24 hour session, bcrypt and a rate limit of five attempts. The trigger phrases are /clarify, "clarify requirements" and "what do I mean by...".

doubt works on the output instead. It forces Claude to re-validate its response when !rv appears in your prompt. The effect is a forced second pass over an answer, on demand, from inside the same conversation.

Together they bracket a single interaction. clarify raises the quality of the instruction before the work starts, and doubt applies pressure to the answer after it arrives. Both are cheap, because neither adds a service, a browser or an account.

The dev plugin runs a four agent decision pipeline

The dev plugin covers developer research and decision-making with two skills, and /tech-decision is the more interesting of them.

The /dev-scan skill searches Reddit, through the Gemini CLI, along with Hacker News, Dev.to and Lobsters, in parallel, then synthesizes consensus, controversies and notable perspectives. The README frames it as a way to understand community sentiment before adopting a tool, and the trigger phrases are of that shape: "developer reactions to...", "what do devs think about...", "A vs B", "which library should I use".

The /tech-decision skill is a multi-phase workflow with four specialised agents running in parallel. The first phase gathers information from four directions at once: a codebase explorer, a docs researcher, the community scan, and the agent council for AI perspectives. Those feed a tradeoff analyzer, which feeds a decision synthesizer that produces an executive summary first, with scored comparisons after.

The README's diagram for it is worth reading, because the phase structure is the point: the four gathering agents are independent, and the two analysis steps are sequential.

Phase 1: Parallel Information Gathering
┌─────────────────┬─────────────────┬─────────────────┬─────────────────┐
│ codebase-       │ docs-           │ dev-scan        │ agent-council   │
│ explorer        │ researcher      │ (community)     │ (AI experts)    │
└────────┬────────┴────────┬────────┴────────┬────────┴────────┬────────┘
         └─────────────────┴─────────────────┴─────────────────┘
                                    │
Phase 2: Analysis & Synthesis       ▼
┌─────────────────────────────────────────────────────────────────────────┐
│                        tradeoff-analyzer                                 │
└─────────────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────────────┐
│                       decision-synthesizer                               │
│                    (Executive Summary First)                             │
└─────────────────────────────────────────────────────────────────────────┘

The cost of this pipeline is that it multiplies model calls and tool access. Four agents in parallel, each doing a different kind of research, is a lot of work for one decision, and the community scan depends on the Gemini CLI being present and configured.

The plugins that reach your accounts and your machine

Half the catalogue touches something outside the editor, and those need a different kind of scrutiny.

gmail is a multi-account Gmail integration with email reading, searching, sending and management. google-calendar is a multi-account calendar integration with parallel querying and conflict detection, so it can look at several calendars at once and tell you when they overlap. kakaotalk sends and reads KakaoTalk messages on macOS using the Accessibility API, which means it drives another application's interface rather than calling a service, and is therefore tied to macOS by construction.

say-summary speaks a short summary of Claude's response using macOS text to speech, in Korean or English, so it is macOS only as well. fetch-tweet fetches full tweet text, author information and engagement data from X/Twitter URLs without authentication, which is the point of it and also the reason to read how it is implemented before you rely on it.

Then there are the media plugins. youtube-digest summarises YouTube videos with a transcript, insights, Korean translation and quizzes. podcast generates Korean podcast episodes from any source using OpenAI text to speech with YouTube auto upload, which is the most outward facing plugin in the set because it publishes to a video platform.

Two are about the session rather than the world. session-wrap is a session wrap up, history analysis and session validation toolkit, and interactive-review is an interactive markdown review with a web UI for visual plan and document approval, which means it starts a local web server to show you the document.

The pattern across these is that each one needs credentials, a local binary, or a platform, and the README does not detail those requirements in the sections shown. The team-assemble plugin is the odd one out, since it dynamically assembles expert agent teams using Claude Code's own agent teams feature, so it needs nothing beyond Claude Code itself.

No releases means the plugin set moves under you

The maintenance picture is thin, and the facts are the facts.

The repository is not archived. The last push was on 2026-04-20, roughly six months before the date of this review, and there are no GitHub releases, so there is no tag to pin, no changelog to diff and no version number in the marketplace entry to check compatibility against.

That matters more here than it would for a library. A library is usually pinned in a lockfile and the version you audited is the version that runs. A Claude Code plugin is installed by name, and the behaviour of a plugin can change when its instructions change, without a version bump that you would notice. For a plugin that rewrites how your assistant answers, that is a real operational difference.

The MIT licence is the permissive one, so there is nothing to negotiate about using, modifying or redistributing the code, and the repository is a normal collection of plugin folders that a team could fork and pin internally if they wanted a stable copy.

The primary language of the repository is recorded as Python, which fits a set of plugins that mostly orchestrate command line tools and services rather than ship a library. The licences of the tools they call are a separate matter from the licence of the plugin itself, and the README does not enumerate them.

Where this plugin set is the wrong choice

Four situations argue against installing any of it.

The first is an environment where you cannot run other model CLIs. The three plugins that most clearly justify their existence, agent-council, dev-scan and tech-decision, all assume the Gemini CLI is present, and the first two also assume GPT and Codex. On a machine with one of those locked down, most of the value disappears and you are left with clarify and doubt, which you could equally write as a saved prompt.

The second is a team that has to explain to a security reviewer what runs during a coding session. A plugin set that reads Gmail, sends Calendar invites, posts to KakaoTalk, publishes to YouTube and opens a local web server for reviews is a lot of surface, and the README does not document the permissions each one takes. Reviewing that per plugin is real work.

The third is macOS, Windows or Linux specifically. The KakaoTalk plugin needs the macOS Accessibility API, say-summary needs macOS text to speech, and the calendar and Gmail integrations are, in practice, most useful to someone already inside those accounts. The set is oriented towards one workflow and one region.

The fourth is a team that wants a stable, reviewable configuration. With no releases and a six month old last push, the plugin set is whatever the branch currently contains, and the plugin instructions are what actually steer the assistant's behaviour, so a change there is a change to how your tooling works.

None of that makes the repository bad. It makes it a collection of experiments and utilities that you adopt individually, pin yourself if it matters, and check the permissions on before you let one near your accounts.

Editorial conclusion

plugins-for-claude-natives fits a Claude Code user who wants several models consulted on one question, a spec interview before coding, or a community scan before choosing a library, since those three plugins work with the tool you already have and add no new service beyond the model CLIs they call. It is the wrong choice if you want a pinned, auditable setup, because the repository publishes no releases and the last push was on 2026-04-20, and it is the wrong choice for the plugins that need your accounts, unless you have read what each one touches. Before installing, add the marketplace, install one plugin at a time, and check the trigger phrases the README lists, since most of these plugins change how Claude responds rather than adding a new command surface.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. README
4. team-attention/plugins-for-claude-natives on GitHub