你的 Skill 散落在几个地方?
Cloud Native
当前 AI Coding 的发展正处在百花齐放的时代,没有永远的王者。模型越来越强,Cursor、Claude Code、Codex 轮番成为阶段性首选;再加上额度限制、响应延迟等现实问题,开发者早就习惯了“鸡蛋不放在一个篮子里”。
除了编程工具,各种通用 Agent 也在不断涌现,每出一个新的,总有人第一时间去尝鲜。
如今,同时使用多个 AI Agent 干活,一边试用刚冒出来的新工具,一边防着正在用的突然掉链子,已经成为这个时代的开发新常态。
然而,工具可以无缝切换,Skill 却无法自动跟随。在 Codex 中更新过的 Skill,Claude Code 里仍是旧版,Cursor 目录下还可能并存一份同名但内容迥异的副本。起初手动复制尚可忍受,时间一长便陷入混乱:哪份才是最新?该用谁覆盖谁?接入新工具时是否还要重复搬运?
这种碎片化的版本管理不仅降低工作效率,更在反复确认中不断消耗人的心力。
一份 Skill,一个中心仓库
Cloud Native
其实社区早就意识到了这个问题,也涌现了不少解法。有人用 Git submodule 或 Monorepo 统一管理,但改个 Prompt 还要走 commit/push/pull 流程,对日常 Coding 来说太重;有人试过 Syncthing 这类通用文件同步工具,但它只懂文件不懂 Skill 语义,双向修改时极易产生覆盖灾难;还有人转向 LangSmith 等 Prompt 管理平台,却发现它们是为 LLM 应用开发设计的,根本不管本地 Agent 配置文件的死活。
开源社区也出现了一些专门针对 Agent Skill 管理的尝试,有的提供了基础的分类检索与统一索引,让散落的 Prompt 有了归处;有的则聚焦跨项目共享,试图解决同一个 Skill 在多个工程中重复维护的问题。
这些探索都精准切中了痛点的某个侧面,但试了一圈下来,我发现它们要么解决了“存储”,没解决“自动分发”;要么解决了“共享”,没解决“多 Agent 目录的实时同步与冲突感知”。当你的 Cursor、Claude Code、Codex 同时开着,需要的是改一处就全部生效、且随时知道哪份是最新的——这个“最后一公里”的闭环,依然没有现成的轮子能完整补上。
既然现有方案都在“部分解决”,那就自己造一个能彻底闭环的。
这就是 Nacos Skill Sync 的由来。它做的事情可以用一句话概括:把 Skill 收敛到一个中心仓库,再按需分发给各个 Agent。
具体来说:
- 只维护一份中心仓库,统一存放 Skill 内容和同步状态;
- 默认用软链接让各 Agent 目录指向中心仓库——改一处,全部生效。环境不支持软链接的话,也可以切换到复制模式,由 CLI 负责把内容同步过去;
- 同步状态随时可查:远端有没有更新、本地有没有改动、是否存在冲突,一眼就知道下一步该干什么。
它提供两种模式,可以按需选择:
模式 | 一句话概括 | 适合谁 |
Registry mode | Nacos AI Registry + 可视化管理 + 跨设备共享 | 想系统管理自己 Skill 库的人,或者需要团队共建 Skill 的场景 |
Local mode | 本地中心仓库 + 软链接/复制同步 + 零服务依赖 | 只想先把本机多个 Agent 的 Skill 统一起来,不想折腾部署 |
两种模式共用一套操作习惯,Local mode 适合先在本机收拢 Skill;Registry mode 则会在对应 profile 下维护独立的 Skill repo 和同步状态。这样不同 profile 的 Skill 不会互相覆盖;真正切换同步来源时,CLI 会先把旧的 profile 的软链接安全落回各 Agent 的本地副本。
先从 Local mode 用起完全没问题,后续需要跨设备共享、团队协作或版本治理时,再切到 Registry mode。
▍Registry mode:远端 Registry 统一管理
如果你手上有十几个 Skill 在持续迭代,或者团队里几个人都在用同一套 Skill,把它们散放在各自的本地目录里,迟早会出问题——版本不一致、改了没人同步、新成员不知道去哪找。
Registry mode 做的,是把 Skill 放进 Nacos AI Registry 这个统一入口:
- 可视化:控制台里浏览、搜索、查看 Skill,不用再翻目录;
- 版本治理:草稿、审核、发布、回滚、label——个人能追踪每个 Skill 的状态,团队能把高频 Skill 沉淀成稳定版本;
- 跨设备同步:家里电脑和公司电脑从同一个 Registry 拉 Skill,新设备接入不用再手动整理;
- 双向流通:远端的可以同步到本地,本地沉淀的也能推上去,不会只留在某一个人的电脑上。
一句话:Skill 从“本地配置文件”变成了可追溯、可共享、可管理的 Registry 资产。
▍Local mode:把本机 Skill 收拢成一份
Local mode 更轻量——无需接入 Nacos AI Registry,仅在本地构建中心仓库,并默认以软链接方式关联各 Agent 的 Skill 目录。
其核心目标是消除多 Agent 环境下的 Skill 冗余与版本不一致问题。当同一份 Skill 在 Codex、Claude、Qoder、Cursor、Kiro、Lingma 等工具中存在多个副本时,Local mode 将其收敛为单一信源,确保修改一处即可全局生效。
目前会自动发现 Codex、Claude、Qoder、QoderWork、Cursor、Kiro、Lingma、CoPaw、OpenClaw,以及通用的 ~/.agents/skills 和 ~/.skills;其他目录也可以通过 skill-sync agent add 手动加入。
已经散落在各 Agent 目录里的 Skill 也可以逐步纳入中心仓库。处理单个 Skill 时,执行 skill-sync add
反过来,如果某个 Skill 不想再由 Skill Sync 管理,执行 skill-sync remove
Local mode 也可以自然升级到 Registry mode。将来想升级到跨设备或团队协作,不用重新整理,绑定一个 profile 就能接入远端。
快速上手
Cloud Native
▍方式一:Agent 自驱动(推荐)
我们提供了一份标准的 SKILL.md,其中完整描述了 nacos-skill-sync 的使用规范与交互流程。只需将下方指令发送给 Agent,它便会自动读取该 Skill、检查本地环境并执行同步;遇到模式选择或内容冲突等需人工决策的节点时,会主动暂停并向你确认。
阅读下面的 Skill,使用它来同步我本地的 Skill:https://nacos.io/skill-sync/SKILL.md
▍方式二:CLI 手动操作
若希望自行控制同步流程,可通过 CLI 完成。
Registry mode(跨设备 / 团队协作)
# 1. 准备 CLI(二选一)curl -fsSL https://nacos.io/nacos-installer.sh | bash -s -- --cli# 或者直接使用 npxnpx @nacos-group/cli@latest skill-sync --help# 2. 配置 CLI profilenpx @nacos-group/cli@latest profile edit test# 3. 添加 Skill,启动同步,查看状态npx @nacos-group/cli@latest --profile test skill-sync add pdfnpx @nacos-group/cli@latest --profile test skill-sync startnpx @nacos-group/cli@latest --profile test skill-sync status
Local mode(单机轻量同步,无需 profile)
npx @nacos-group/cli skill-sync add pdfnpx @nacos-group/cli skill-sync startnpx @nacos-group/cli skill-sync status
使用案例
Cloud Native
下面两个例子分别对应 Local mode 和 Registry mode:一个从个人本机开始,把高频工作流先收拢起来;另一个面向跨设备和团队协作,把共同规范放进 Registry。
▍案例一:记录工作内容&生成周报 Skill
很多人每天都会在多个 Agent 里处理不同任务:用 Codex 改代码,用 Claude Code 查日志,用 Cursor 补测试。任务做完后,工作记录也跟着散在不同对话里。到周五再回头整理,常常要翻聊天、补背景、合并重复事项,最后写出来的周报还容易变成流水账。
这个问题可以先从本机解决:把每天怎么记录、周报怎么合并这些习惯沉淀成一个 biweekly-work-report Skill,放在 Local mode 里给本机 Agent 共用。任务结束时记录结果和影响,遇到同主题工作时合并更新,生成周报时保留进展、风险和下周计划。
使用 Skill Sync 将某个 Agent 创建的 biweekly-work-report同步至所有的 Agent:
后续所有 Agent 都可以使用这个 Skill:
Codex:
Claude Code:
这样做不需要先准备服务端,也不用把私人工作记录同步出去。更关键的是,这不是把 Skill 复制一次就结束:后面在 Codex 里调整了周报字段,Claude Code 和 Cursor 看到的也会是更新后的 work-report。
▍案例二:文档统一格式 Skill
团队文档的问题更典型。接口文档、技术方案、故障复盘可能由不同人、不同设备、不同 Agent 生成,标题层级、参数表字段、风险说明和评审清单很容易各写各的。靠口头提醒不稳定,复制模板又经常复制到旧版本。
这类格式规范更适合放进 Registry mode。团队可以把技术方案结构、API 参数表字段顺序、术语写法、截图说明和评审检查项整理成一个 doc-format Skill。
使用 Skill Sync 将 AI Registry 中的 doc-format同步至所有多个设备的 Agent:
后续不同设备的 Agent 都可以使用这个 Skill:
设备 A Codex:
设备 B Qoder:
规范更新后,只维护 Registry 里的这一份。公司电脑、家里电脑、新设备,以及团队成员的 Agent 同步后都会继续使用同一套格式。
日常维护:看 status 就够了
Cloud Native
用起来之后,日常只看一个命令的输出:
Mode: nacosProfile: teamSync daemon: running (pid: 12345)SKILL STATUS AGENTS NEXTpdf Synced codex,claude,qoder -review Local changes codex,claude,qoder auto-upload pendingdraft Uploaded codex,claude,qoder waiting publishblock Upload blocked codex,claude,qoder Nacos draft exists; review/clear it, auto-upload will retrytriage Conflict codex,claude,qoder skill-sync resolve triage
每个 Skill 什么状态、覆盖了哪些 Agent、下一步该干什么——都在这了。
状态 | 意味着什么 |
| 一切正常,不用管 |
| Local mode 下已经链接到本地中心仓库 |
| 本地有改动;开启 auto-upload 时 daemon 会自动上传,关闭时会保留在本地 |
| 已上传为草稿,等待审核或发布 |
| 本地和远端都改了,需要你选一个版本 |
| 远端已有草稿或审核中版本,需要先在 Nacos 里处理;处理后会自动重试 |
遇到冲突也不用慌,resolve命令可以选以远端 Registry、中心仓库还是某个 Agent 的版本为准。策略默认保守——不会擅自替你做选择,你明确指定了来源才会执行。
让 Skill 有一份可信来源
Cloud Native
多 Agent 并行会越来越常见,真正需要管理的,不只是用哪个 Agent,而是这些 Agent 共同依赖的 Skill。
更理想的方式,是把 Skill 放进 Registry:有统一入口,有版本记录,也能在多台设备和团队成员之间共享。Local mode 则提供了一个更轻的入口,不需要先准备服务端,也能先把本机散落的 Skill 收拢起来。
无论从哪种方式开始,目标都是一样的:让 Skill 有一份可信来源。Agent 可以换,Skill 不应该跟着散。
欢迎大家使用 Nacos Skill Sync。如果在使用中发现任何问题,或有更好的改进建议,也欢迎随时反馈给我们。
相关链接:
- Nacos 开源地址:github.com/alibaba/nacos
- Nacos CLI:https://github.com/nacos-group/nacos-cli
- Nacos Skill Sync:https://nacos.io/skill-sync/SKILL.md
Your Skills are scattered across several places?
Cloud Native
The current development of AI coding tools is in a flourishing era with no permanent king. Models keep getting stronger, with Cursor, Claude Code, and Codex rotating as the top picks at different stages. Coupled with real-world issues like usage limits and response latency, developers have long gotten used to "not putting all their eggs in one basket."
Beyond coding tools, all kinds of general-purpose Agents keep emerging, and whenever a new one appears, someone is always first to try it.
Nowadays, using multiple AI Agents at once—trying out newly released tools while guarding against the ones you rely on suddenly failing—has become the new normal in development.
However, tools can switch seamlessly, but Skills cannot follow automatically. A Skill updated in Codex is still the old version in Claude Code, and the Cursor directory may even hold a same-named copy with completely different content. At first, manually copying is tolerable, but over time it descends into chaos: which copy is the latest? Which one should overwrite which? Do you have to keep re-copying when integrating a new tool?
This fragmented version management not only reduces work efficiency but also constantly drains people's mental energy through repeated checking.
One Skill, one central repository
Cloud Native
The community actually recognized this problem long ago, and many solutions have emerged. Some use Git submodule or Monorepo for unified management, but changing a Prompt still requires going through commit/push/pull, which is too heavy for everyday coding; some tried general-purpose file sync tools like Syncthing, but they only understand files, not Skill semantics, making overwrite disasters highly likely during bidirectional modifications; others turned to Prompt management platforms like LangSmith, only to find they are designed for LLM application development and don't care about local Agent configuration files at all.
The open source community has also seen attempts specifically aimed at Agent Skill management. Some provide basic categorization, search, and unified indexing, giving scattered Prompts a place to belong; others focus on cross-project sharing, trying to solve the problem of maintaining the same Skill repeatedly across multiple projects.
These explorations all precisely hit one side of the pain point, but after trying them all, I found they either solved "storage" without solving "automatic distribution," or solved "sharing" without solving "real-time synchronization and conflict awareness across multiple Agent directories." When your Cursor, Claude Code, and Codex are all open at the same time, what you need is for one change to take effect everywhere and to always know which copy is the latest—this "last mile" closed loop still has no ready-made solution to fully fill.
Since existing solutions only "partially solve" the problem, I decided to build one that can completely close the loop.
That is the origin of Nacos Skill Sync. What it does can be summarized in one sentence: converge Skills into a central repository, then distribute them to each Agent as needed.
Specifically:
Maintain only one central repository, which uniformly stores Skill content and sync status;
By default, use symbolic links to point each Agent directory to the central repository—change once, take effect everywhere. If the environment does not support symbolic links, you can also switch to copy mode, where the CLI handles syncing the content over;
Sync status is always queryable: whether the remote has updates, whether local changes exist, whether there is a conflict—you can tell at a glance what to do next.
It provides two modes, and you can choose as needed:
Mode
One-sentence summary
Who it suits
Registry mode
Nacos AI Registry + visual management + cross-device sharing
People who want to systematically manage their Skill library, or scenarios where a team needs to co-build Skills
Local mode
Local central repository + symbolic link/copy sync + zero service dependencies
For those who just want to unify Skills across multiple Agents on one machine without dealing with deployment
Both modes share the same operating habits. Local mode is suitable for first consolidating Skills on the local machine; Registry mode maintains an independent Skill repo and sync status under the corresponding profile. This way, Skills from different profiles do not overwrite each other; when actually switching sync sources, the CLI first safely returns the old profile's symbolic links to each Agent's local copies.
It is completely fine to start with Local mode, and switch to Registry mode later when you need cross-device sharing, team collaboration, or version governance.
▍ Registry mode: unified management via remote Registry
If you have a dozen or so Skills continuously iterating, or several people on the team all using the same set of Skills, leaving them scattered in separate local directories will eventually cause problems—version inconsistency, changes nobody syncs, and new members not knowing where to find them.
What Registry mode does is put Skills into Nacos AI Registry, a unified entry point:
Visualization: browse, search, and view Skills in the console without digging through directories;
Version governance: draft, review, publish, rollback, label—individuals can track the status of each Skill, and teams can consolidate frequently used Skills into stable versions;
Cross-device synchronization: pull Skills from the same Registry on home and work computers, so new devices don't need manual organization;
Bidirectional flow: remote Skills can sync to local, and locally accumulated ones can be pushed up, so they don't stay on just one person's computer.
In one sentence: Skills go from "local configuration files" to traceable, shareable, manageable Registry assets.
▍ Local mode: consolidate local Skills into one copy
Local mode is lighter—no need to connect to Nacos AI Registry. It only builds a central repository locally and, by default, associates each Agent's Skill directory via symbolic links.
Its core goal is to eliminate Skill redundancy and version inconsistency in multi-Agent environments. When the same Skill exists as multiple copies in tools such as Codex, Claude, Qoder, Cursor, Kiro, Lingma, etc., Local mode converges them into a single source of truth, ensuring that modifying one place takes effect everywhere.
Currently, it automatically discovers Codex, Claude, Qoder, QoderWork, Cursor, Kiro, Lingma, CoPaw, OpenClaw, as well as the generic ~/.agents/skills and ~/.skills; other directories can also be added manually via skill-sync agent add.
Skills already scattered across Agent directories can also be gradually brought into the central repository. When handling a single Skill, run skill-sync add
Conversely, if you no longer want a Skill managed by Skill Sync, run skill-sync remove
Local mode can also naturally upgrade to Registry mode. If you want to upgrade to cross-device or team collaboration later, you don't need to reorganize anything—just bind a profile to connect to the remote.
Quick start
Cloud Native
▍ Method 1: Agent self-driven (recommended)
We provide a standard SKILL.md that fully describes the usage conventions and interaction flow of nacos-skill-sync. Simply send the instructions below to the Agent, and it will automatically read the Skill, check the local environment, and perform the sync; at decision points requiring human input, such as mode selection or content conflicts, it will proactively pause and ask for your confirmation.
▍ Method 2: CLI manual operation
If you prefer to control the sync process yourself, you can do it through the CLI.
Registry mode (cross-device / team collaboration)
Local mode (single-machine lightweight sync, no profile needed)
Use cases
Cloud Native
The two examples below correspond to Local mode and Registry mode: one starts from a personal machine and consolidates high-frequency workflows first; the other targets cross-device and team collaboration, putting shared conventions into the Registry.
▍ Case 1: work logging & weekly report Skill
Many people handle different tasks across multiple Agents every day: fixing code with Codex, checking logs with Claude Code, and writing tests with Cursor. After tasks are done, work records end up scattered across different conversations. By Friday, when you look back to organize, you often have to search chats, fill in background, and merge duplicate items—and the weekly report you finally write tends to become a plain log.
This problem can be solved locally first: turn your daily logging and weekly report merging habits into a biweekly-work-report Skill and put it in Local mode for all local Agents to share. When a task ends, record the result and impact; when related work appears, merge and update; when generating the weekly report, keep progress, risks, and next week's plan.
Use Skill Sync to sync the biweekly-work-report created by one Agent to all Agents:
From then on, all Agents can use this Skill:
Codex:
Claude Code:
This approach doesn't require preparing a server first, and it doesn't require syncing private work records outward. More importantly, this isn't just copying the Skill once and being done: later, if you adjust the weekly report fields in Codex, Claude Code and Cursor will also see the updated work-report.
▍ Case 2: unified document format Skill
Team documentation problems are even more typical. Interface docs, technical proposals, and fault retrospectives may be generated by different people, different devices, and different Agents, and heading levels, parameter table fields, risk descriptions, and review checklists can easily be written differently by each person. Relying on verbal reminders is unstable, and copying templates often results in copying old versions.
This kind of format convention is better suited for Registry mode. The team can organize technical proposal structure, API parameter table field order, terminology conventions, screenshot explanations, and review checklist items into a doc-format Skill.
Use Skill Sync to sync doc-format from AI Registry to Agents across multiple devices:
From then on, Agents on different devices can all use this Skill:
Device A Codex:
Device B Qoder:
After the convention is updated, only maintain this one copy in the Registry. Company computers, home computers, new devices, and team members' Agents will all continue using the same format after syncing.
Daily maintenance: just look at status
Cloud Native
Once you start using it, daily maintenance only requires looking at the output of one command:
What status each Skill is in, which Agents it covers, and what to do next—it's all here.
Status
What it means
Synced
Everything is normal, nothing to do
Linked
Already linked to the local central repository in Local mode
Local changes
There are local changes; when auto-upload is enabled, the daemon uploads them automatically; when disabled, they remain local
Uploaded
Uploaded as draft, waiting for review or publication
Conflict
Both local and remote changed; you need to choose one version
Upload blocked
The remote already has a draft or version under review; you need to handle it in Nacos first; after handling, it will retry automatically
When you encounter a conflict, don't panic—the resolve command lets you choose whether to use the remote Registry, central repository, or a specific Agent's version as the standard. The strategy is conservative by default—it won't make choices for you on its own; it only executes when you explicitly specify a source.
Give Skills a trusted source
Cloud Native
Multi-Agent parallelism will become increasingly common. What truly needs managing isn't just which Agent to use, but the Skills these Agents jointly depend on.
A more ideal approach is to put Skills into a Registry: with a unified entry point, version history, and the ability to share across multiple devices and team members. Local mode offers a lighter entry point—no need to prepare a server first, and you can still consolidate scattered local Skills.
No matter which approach you start with, the goal is the same: give Skills a trusted source. Agents can be switched, but Skills shouldn't remain scattered.
You are welcome to use Nacos Skill Sync. If you encounter any issues during use, or have better improvement suggestions, feel free to give us feedback at any time.
Related links:
Nacos open source address: github.com/alibaba/nacos
Nacos CLI: https://github.com/nacos-group/nacos-cli
Nacos Skill Sync: https://nacos.io/skill-sync/SKILL.md









