Cursor's plugin spec and official plugin catalog for agent skills
A Cursor-maintained plugin specification plus a catalog of official plugins. Each plugin is described by a plugin.json manifest (name, author, category) and bundles skills, commands, scripts, or connectors into units an agent can invoke directly. The collection splits into two halves: developer-workflow plugins (PR review canvases, parallel subagent orchestration, CLI patterns agents can run reliably, scaffolding and validation via create-plugin) and integrations that expose external systems such as GitHub, Playwright, BigQuery, Salesforce and Gmail to the agent. Useful if you are packaging reusable capabilities for coding agents, or want a reference for how the vendor structures and distributes them.
Cursor's plugin catalog is two things at once: a spec for packaging agent skills, and a browsable index of the ones that already exist. A plugin is a small, self-describing unit — slug, display name, author, category, and a single line stating what it does — that an agent picks up when a task calls for it. The presence of create-plugin, which scaffolds and validates new entries, is the clearest signal that this is a governed format rather than a folder of loose prompts. Three audiences are served at once: developers who want an agent to follow a review or shipping process, teams who need the agent to reach GitHub, Salesforce or Zoom, and people who simply want a learning loop sitting on top of their editor.
Cursor 的插件目录同时是两个东西:一套把 agent skill 打包起来的规范,以及一份可浏览的现有插件索引。一个插件是一个自描述的小单元——slug、显示名、作者、分类,再加一句话说明它是干什么的——agent 在任务需要时把它加载进来。目录里有 create-plugin 这个负责生成骨架并校验新插件的 skill,这本身就是最明确的信号:这是一套有约束的格式,不是一堆散落的 prompt。它同时服务三类人:想让 agent 按固定流程做代码评审和发布的开发者,需要 agent 去读写 GitHub、Salesforce、Zoom 这类外部系统的团队,以及只是想在编辑器里加一个学习闭环的人。
Agent skills have a distribution problem. The useful ones live in personal dotfiles, in a gist someone linked on social media, or buried in one team's repo, and there is no shared way to answer even basic questions: does a skill for this already exist, who wrote it, what does it actually do, is it still maintained? The catalog answers the discovery half; the spec answers the packaging half. Concretely, a spec buys you a stable identity per plugin, attributable authorship so a name in the listing maps back to a real plugin.json, and a category so you can narrow the search before reading anything. Without that, "just write a skill" degrades into every repository inventing its own layout and vocabulary, and nothing composes.
agent skill 有个分发难题。好用的那些要么躺在个人 dotfiles 里,要么是某人在社交媒体上贴的一段 gist,要么埋在某个团队的仓库深处;连最基本的问题都没有统一答案:这件事是不是已经有人写过了?谁写的?它到底做什么?还在维护吗?目录解决了「找」的一半,规范解决了「打包」的一半。具体来说,规范给你三样东西:每个插件有稳定身份;作者可追溯,目录里署的名字能对回真实的 plugin.json;有分类,让你在读内容之前先把范围收窄。没有这些,「随手写个 skill」就会退化成每个仓库自造一套目录结构和叫法,互相之间接不上。
The catalog sorts entries into four categories. Developer Tools is the most opinionated group: process skills for review, verification, memory, parallel execution and shipping. Integrations is by far the largest — vendor connectors that let an agent read and act inside an external product, from GitHub and Playwright to Salesforce, Workday and a long tail of ad, CRM and finance platforms. Productivity gathers the everyday work surfaces: Gmail, Google Drive, Docs, Sheets, Slides, Calendar, plus Microsoft Teams and SharePoint. Utilities is the smallest bucket here, holding the single Teaching plugin for skill mapping, practice plans and learning retrospectives. The split is practical rather than academic: it tells you whether a plugin changes how the agent works, what it can reach, or where it sits in your day.
目录把插件分成四类。Developer Tools 是最有主张的一类:评审、验证、记忆、并行执行、发布这些流程型 skill 都在这里。Integrations 数量最多,是各类厂商连接器,让 agent 能读写外部产品——从 GitHub、Playwright 一直到 Salesforce、Workday,以及一大串广告、CRM、财务平台。Productivity 收的是日常工作面:Gmail、Google Drive、Docs、Sheets、Slides、Calendar,加上 Microsoft Teams 和 SharePoint。Utilities 是最小的一格,只有 Teaching 一个插件,做技能映射、练习计划和复盘。这个划分是实用导向而非学术分类:它告诉你的其实是——这个插件改变的是 agent 的工作方式、它能触达的系统,还是它在你这天里的位置。
Read the Developer Tools entries together and a handful of themes emerge. Verification over generation: thermos runs a thermo-nuclear branch review with security and correctness audits plus a harsh code-quality rubric, pr-review-canvas renders diffs grouped by importance, docs-canvas makes documentation navigable. Parallelism with structure: orchestrate fans a large task out across planner, worker and verifier agents with explicit handoffs; pstack is built around workflows you can parallelize with confidence. Memory: continual-learning appends high-signal bullets to AGENTS.md from transcripts instead of rewriting it wholesale. Self-checking loops: ralph-loop does iterative self-referential loops, and grok-voice ships a log-driven fix loop for voice sessions. Escalation: advisor says consult a stronger model before big decisions and before declaring done. And reality checks: agent-compatibility audits startup, validation and docs against what the repo actually does; cli-for-agent collects patterns for CLIs an agent can run reliably — flags, examples in help, pipelines, errors, idempotency, dry-run.
把 Developer Tools 这一类连着读,几个主题会浮出来。验证重于生成:thermos 做「热核级」分支评审,带安全与正确性审计、用很苛刻的代码质量评分表;pr-review-canvas 把 diff 按重要程度分组渲染;docs-canvas 把文档变成可导航的画布。有结构的并行:orchestrate 把大任务拆给 planner、worker、verifier 三类 agent,交接有明确格式;pstack 的核心就是让你敢并行。记忆:continual-learning 从对话记录里提炼高信号要点、增量追加到 AGENTS.md,而不是整篇重写。自检循环:ralph-loop 做自引用的迭代循环,grok-voice 给语音会话配了一个靠日志驱动的修复循环。升级求助:advisor 主张在重大决策之前、在宣布完成之前,先去问更强的模型。还有对照现实:agent-compatibility 拿真实仓库去校验启动、验证和文档是否还对得上;cli-for-agent 收集的是「让 agent 能可靠跑起来」的 CLI 设计模式——flag、带示例的 help、管道、错误、幂等、dry-run。
The Integrations list is where the catalog's ambition shows. Developer infrastructure (GitHub, Playwright, BigQuery, Databricks Genie, Statsig, Vanta), the whole Google Workspace and Microsoft surface, CRM and sales tooling (Salesforce, HubSpot, Attio, Zoho, Outreach, Amplemarket), meeting intelligence (Zoom, Gong, Circleback, Fireflies, Otter, Fathom), marketing and ads (Google Ads, Meta Ads, TikTok for Business, X Ads, Klaviyo, Customer.io, Semrush, Ahrefs), HR and payroll (Workday, Gusto, Rippling, Ashby, Workable), commerce and ops (Shopify, Square, Jobber, Gopuff, Navan, Brex, Xero, QuickBooks), and a dense cluster of brokerage and fintech skills. Two design threads are worth noting. First, write-capable skills are careful about confirmation: Google Ads makes "previewed changes", QuickBooks requires a clear preview and your confirmation before anything is sent, Tesla sends "commands you confirm", Coinbase and Robinhood preview before placing a trade. Second, many of these are thin layers over a vendor's own MCP server or API — Gopuff says so explicitly, Tailscale exists purely to reach MCP servers on your private tailnet. A few descriptions are even written in the voice of Grok, which suggests the catalog feeds more than one assistant surface.
Integrations 这一类最能看出目录的野心。开发者基础设施(GitHub、Playwright、BigQuery、Databricks Genie、Statsig、Vanta)、整片 Google Workspace 与 Microsoft 表面、CRM 与销售工具(Salesforce、HubSpot、Attio、Zoho、Outreach、Amplemarket)、会议智能(Zoom、Gong、Circleback、Fireflies、Otter、Fathom)、营销与广告(Google Ads、Meta Ads、TikTok for Business、X Ads、Klaviyo、Customer.io、Semrush、Ahrefs)、人事与薪酬(Workday、Gusto、Rippling、Ashby、Workable)、电商与运营(Shopify、Square、Jobber、Gopuff、Navan、Brex、Xero、QuickBooks),还有一整片券商与金融科技。有两条设计线索值得单独拎出来。一是写入类 skill 对确认很谨慎:Google Ads 只做「预览后的变更」,QuickBooks 要求先给出清晰预览、经你确认才发送,Tesla 发的是「你确认过的命令」,Coinbase 和 Robinhood 都是先预览再下单。二是不少插件其实是厂商 MCP server 或 API 的薄壳:Gopuff 明说了这点,Tailscale 存在的意义就是让你的私有 tailnet 里的 MCP server 可达。另有几条描述直接用 Grok 的口吻写,说明这份目录服务的并不止一个助手入口。
Every plugin carries a plugin.json manifest, and the author shown in the catalog is required to match that manifest's author.name — Cursor's own plugins list plugins@cursor.com there. This is a small rule with real consequences: attribution is checkable rather than decorative, so you can tell a first-party skill from a community one at a glance and know which name to search for when something misbehaves. The third-party authors visible in the listing are few and identifiable: Eric Zakariasson (continual-learning, cursor-team-kit, cli-for-agent, grok-voice), Lauren Tan (pstack) and Dylan Gattey (dyl-stack). The last one is built explicitly "on top of pstack", which is the catalog quietly demonstrating that skills stack: one author's workflow discipline can be a dependency of another's style.
每个插件都带一份 plugin.json 清单,目录里展示的作者必须与清单里的 author.name 一致——Cursor 自家的插件在那里填的是 plugins@cursor.com。这条规则很小,但后果实在:署名是可核对的,不是装饰,你一眼就能分辨官方 skill 和社区 skill,出问题时也知道该搜谁的名字。目录里出现的第三方作者不多且好认:Eric Zakariasson(continual-learning、cursor-team-kit、cli-for-agent、grok-voice)、Lauren Tan(pstack)、Dylan Gattey(dyl-stack)。其中 dyl-stack 明写「建在 pstack 之上」,这其实是目录在不动声色地示范 skill 是可以叠加的:一个人的流程纪律,可以是另一个人风格的前置依赖。
The path to publishing starts with create-plugin, which scaffolds a new plugin and validates it against the spec — so you find out early whether your manifest and layout are acceptable instead of after you have written the whole thing. If what you are actually building is an app or an automation rather than a skill, cursor-sdk covers the TypeScript route. Two other plugins double as authoring guidance: cli-for-agent tells you how to shape a CLI so an agent can drive it without flailing (clear flags, help text with examples, sane error output, idempotent operations, a dry-run mode), and agent-compatibility is what you want pointed at your own repo when the docs and the startup path have drifted apart. Then fill in author.name correctly and the listing will credit you.
要发布一个插件,起点是 create-plugin:它生成骨架,并按规范做校验——所以在还没写完之前你就知道清单和结构过不过得了,而不是全写完再返工。如果你要做的其实是应用或自动化而不是 skill,那就走 cursor-sdk 这条 TypeScript 路线。另外两个插件兼作写作指南:cli-for-agent 讲怎么把 CLI 设计成 agent 能顺利驱动的样子(清晰的 flag、带示例的 help、像样的错误输出、幂等操作、dry-run 模式);agent-compatibility 则适合对准你自己的仓库——尤其是当文档和实际启动路径已经对不上的时候。最后把 author.name 填对,目录自然会署你的名。
A few honest limits. The catalog is Cursor-shaped: skills are written against this runtime's affordances, so lifting one into a different agent framework means rework, not copy-paste. Integrations inherit every weakness of what they wrap — if the vendor's API or MCP server is down, rate-limited, or the OAuth scopes are wrong, the skill cannot save you, and the failure messages will come from someone else's system. Some entries are explicitly marked Beta (Salesforce Headless 360), which should set expectations for rough edges. Skills that move money or send mail on your behalf deserve real caution: the preview-and-confirm pattern is reassuring, but confirmation is only as good as the person clicking, and agentic brokerage accounts are still brokerage accounts. Finally, the spec guarantees format, not quality — a first-party skill and a community skill look identical in the listing, and several developer-tooling entries overlap heavily, so choosing among thermos, orchestrate, pstack and dyl-stack is a judgment call about your process rather than a feature comparison.
几条实在的边界。这份目录是 Cursor 形状的:skill 是照着这套运行时的能力写的,搬到别的 agent 框架里是重写而不是复制粘贴。集成类插件会继承它们所包裹对象的一切弱点——厂商 API 挂了、被限流、OAuth scope 配错了,skill 救不了你,报错还是别人系统给的。部分条目明确标了 Beta(Salesforce Headless 360),对它该有不成熟的心理预期。会动钱、会代你发信的 skill 需要格外小心:预览加确认的模式让人安心,但确认的质量取决于点确认的那个人,而且 agentic 券商账户说到底还是券商账户。最后,规范保证的是格式而不是质量——官方 skill 和社区 skill 在列表里长得一模一样;开发工具类的若干条目重合度也很高,在 thermos、orchestrate、pstack、dyl-stack 之间选,靠的是对你自身流程的判断,不是比功能表。