面向 AI 编程工具的插件、Agent 与技能库

通过一套插件、Agent、技能（Skills）和命令合集，即可跨 Claude Code、OpenAI Codex CLI、Cursor、OpenCode、Antigravity CLI、GitHub Copilot 以及 Pi 运行。所谓载体（Harness），即加载并运行这些组件的编程工具。插件覆盖了 Python 和 JavaScript 开发、代码审查、测试、基础设施、安全等各领域工作。

该目录共收录 94 个插件，包括 92 个本地插件与 2 个外部条目。本地源码包含 202 个 Agent、185 项技能以及 105 个命令。所有组件共享同一套 Markdown 源码，并针对各个载体的特性自适应调整安装方式与功能支持。

请选择你所使用的载体并按照其配置路径进行操作。可用组件取决于安装路径以及相应载体的功能支持。

有关模型映射、工具权限、Hook 以及安装差异的更多信息，请参阅能力矩阵。

如需安装单个技能，可使用纯技能安装器（skills-only installer）并选择你的编程工具。

选择符合你工作需求的插件：

查阅插件目录可了解全部 94 个插件，查看使用指南可获取命令和示例。

在 Claude Code 内运行以下命令以注册该市场并安装 Python 开发插件：

/plugin marketplace add wshobson/agents /plugin install python-development@claude-code-workflows

随后即可让 Claude 调用 Python Agent 处理任务，或运行该插件的项目脚手架生成命令：

/python-development:python-scaffold Create a FastAPI service with tests

在安装有最新 Codex CLI 的终端中运行以下命令：

codex plugin marketplace add wshobson/agents codex plugin add python-development@claude-code-workflows

Codex 的原生 Manifest 会暴露源码中的技能。生成的 TOML Agent 以及由命令衍生的技能则采用单独的适配器路径。此外，原生安装会读取完整的原始技能正文，而生成的版本则会将超出大小的内容拆分为引用文件，以适配 8 KB 的上限。选择安装路径前，请参阅 Codex 配置详情。

Cursor 使用仓库中已提交的插件市场注册表。请按照 Cursor 配置说明添加市场并安装插件。

你可以使用 GitHub CLI 的 gh skill 命令或 npx skills 安装器安装单个技能。系统提示时选择你的编程工具即可。这两种安装器无需克隆仓库即可直接从 GitHub 读取技能，并且在安装时不会引入插件所属的 Agent、命令或 Hook。

gh skill install wshobson/agents python-testing-patterns npx skills add wshobson/agents --skill python-testing-patterns

gh skill 需要 GitHub CLI 2.90 或更高版本。npx 命令需要 Node.js 和 npm。关于其他 Agent、安装作用域及版本锁定的信息，请参阅纯技能安装指南。

对于 OpenCode、Antigravity CLI、GitHub Copilot 和 Pi，请克隆该仓库并使用对应载体的安装器。需提前安装 uv、Python 3.12 或更高版本，以及你打算使用的编程工具。以下克隆命令使用 GitHub CLI：

gh repo clone wshobson/agents ~/agents cd ~/agents

运行针对对应工具的命令。每个目标都会生成相应构件，并将所有本地插件软链接至该工具的用户配置目录中：

make install-opencode make install-antigravity make install-copilot make install-pi

Pi 的 Agent 需要其参考实现的 subagent 扩展或兼容扩展。有关配置路径与卸载命令，请参阅全局安装指南。

上述 Agent、技能和命令数量统计仅涵盖本地源文件。外部插件提供其自有的组件与安装要求。

一个插件会将相关的 Agent、技能和命令聚合在 plugins/ 目录下。例如，python-development 包含 3 个 Agent、16 项技能以及 1 个脚手架生成命令。所选的载体会根据自身的发现与调用规则来加载这些已安装组件。

贡献者基于 Claude Code 插件格式维护单一的 Markdown 源码。原生注册表在支持的情况下会直接指向该源码，适配器则负责将其生成各载体所需的格式。在此过程中，命令可能会转换为技能或 Prompt 模板，而 Agent 格式、模型选择与工具权限也会因载体不同而有所差异。

源码 Agent 会指定诸如 opus、sonnet、haiku、fable 或 inherit 等模型别名。适配器将这些别名映射至每个载体的模型设置，包括为部分 inherit 条目指定固定默认值。详见模型配置与适配器映射规则。

该仓库会对插件结构、生成的构件以及文档进行校验。贡献者可在仓库根目录下运行以下检查：

make generate-all make validate STRICT=1 make garden

plugin-eval 还提供了无需调用模型的静态技能检查。例如：

uv run --project plugins/plugin-eval plugin-eval score plugins/python-development/skills/python-testing-patterns --depth quick

LLM 裁判（LLM judge）与蒙特卡洛层目前处于实验阶段，尚未经过人类标注数据的验证。其评分仅代表模型评估结果与重复模拟结论，并不等同于任务在实际项目中必然执行成功。这部分功能需要安装可选的 llm 依赖。若配置了 ANTHROPIC_API_KEY，模型调用将使用该 Key 并向对应 API 账户计费；否则将使用你的 Claude Code 登录凭据。有关设置与限制说明，请参阅评估指南。

如需提交修改建议，请参阅 CONTRIBUTING.md；编写插件内容请参阅编写规范。

外部记忆集成

Pensyve 仍保留为 Claude Code 市场中的可选外部条目。Pensyve 由 Major7 Apps 维护，本项目维护者也是 Major7 Apps 的创始人。其开源运行时可运行在本地或自建服务器上。Pensyve Cloud 已于 2026 年 10 月 1 日停止服务，其控制台（Dashboard）、API 以及托管的 MCP 端点均已不可用。配置该运行时请参考自建托管指南。

本项目遵循 MIT 许可协议。外部插件遵循其各自的许可证。

Plugins, agents, and skills for AI coding tools

Use one collection of plugins, agents, skills, and commands across Claude Code, OpenAI Codex CLI, Cursor, OpenCode, Antigravity CLI, GitHub Copilot, and Pi. A harness is the coding tool that loads and runs the components. Plugins cover Python and JavaScript development, code review, testing, infrastructure, security, and other work.

The catalog contains 94 plugins, including 92 local plugins and 2 external entries. The local source contains 202 agents, 185 skills, and 105 commands. The components share a Markdown source, with installation and capabilities adapted to each harness.

Choose the harness you use and follow its setup route. Available components depend on the installation route and the harness's capabilities.

See the capability matrix for model mappings, tool permissions, hooks, and installation differences.

For individual skills, use a skills-only installer and select your coding tool.

Choose a plugin for the work you do:

See the plugin catalog for all 94 plugins and the usage guide for commands and examples.

Run these commands inside Claude Code to register the marketplace and install the Python development plugin:

/plugin marketplace add wshobson/agents
/plugin install python-development@claude-code-workflows

Then ask Claude to use the Python agents for your task, or run the plugin's scaffolding command:

/python-development:python-scaffold Create a FastAPI service with tests

Run these commands in a terminal with a current Codex CLI:

codex plugin marketplace add wshobson/agents
codex plugin add python-development@claude-code-workflows

The native Codex manifests expose source skills. The generated TOML agents and command-derived skills use a separate adapter path. Native installation also reads the original skill bodies, while generated copies split oversized bodies into reference files for the 8 KB budget. See the Codex setup details before choosing an installation route.

Cursor uses the committed plugin marketplace registry. Follow the Cursor setup instructions to add the marketplace and install a plugin.

Install an individual skill with GitHub CLI's gh skill command or the npx skills installer. Select your coding tool when prompted. Both installers read skills from GitHub without cloning the repository and install them without the plugin's agents, commands, or hooks.

gh skill install wshobson/agents python-testing-patterns
npx skills add wshobson/agents --skill python-testing-patterns

gh skill requires GitHub CLI 2.90 or later. The npx command requires Node.js and npm. See skills-only installation for other agents, install scopes, and version pinning.

For OpenCode, Antigravity CLI, GitHub Copilot, and Pi, clone the repository and use the installer for your harness. Install uv, Python 3.12 or later, and the coding tool you want to use. The following clone command uses GitHub CLI:

gh repo clone wshobson/agents ~/agents
cd ~/agents

Run the command for your tool. Each target generates the artifacts and links all local plugins into that tool's user configuration:

make install-opencode
make install-antigravity
make install-copilot
make install-pi

Pi's agents require its reference subagent extension or a compatible extension. See the global install guide for configuration paths and uninstall commands.

The agent, skill, and command counts cover local source files. External plugins provide their own components and installation requirements.

A plugin groups related agents, skills, and commands under plugins/. For example, python-development contains three agents, sixteen skills, and a scaffolding command. The chosen harness loads the installed components according to its own discovery and invocation rules.

Contributors maintain one Markdown source in the Claude Code plugin format. Native registries point to that source where supported, and adapters generate the formats needed by each harness. Commands may become skills or prompt templates, and agent formats, model choices, and tool permissions vary by harness.

Source agents specify model aliases such as opus, sonnet, haiku, fable, or inherit. Adapters map those aliases to each harness's model settings, including fixed defaults for some inherit entries. See model configuration and the adapter mappings.

The repository validates plugin structure, generated artifacts, and documentation. Contributors can run these checks from the repository root:

make generate-all
make validate STRICT=1
make garden

plugin-eval also provides static skill checks without model calls. For example:

uv run --project plugins/plugin-eval plugin-eval score plugins/python-development/skills/python-testing-patterns --depth quick

The LLM judge and Monte Carlo layers are experimental and are not validated against human labels. Their scores describe model assessments and repeated simulations and do not prove that a task succeeds in a project. They require the optional llm dependencies. Model calls use ANTHROPIC_API_KEY when it is set and bill that API account. Otherwise they use your Claude Code login. See the evaluation guide for setup and limits.

See CONTRIBUTING.md to propose a change and authoring conventions to write plugin content.

External memory integration

Pensyve remains an optional external Claude Code marketplace entry. Major7 Apps maintains Pensyve, and this repository's maintainer founded Major7 Apps. Its open-source runtime can run locally or on your own server. Pensyve Cloud closed on October 1, 2026, and its dashboard, API, and hosted MCP endpoints are unavailable. Use the self-hosting guide when configuring the runtime.

The repository is MIT licensed. External plugins have their own licenses.