notenerd2k downloadsManage AI skills, agents, commands, and rules for Claude Code, Cursor, Codex, and more, with usage insights and MCP visibility.
Browse, tag, and organize AI skills, agents, commands, rules, agent memories, and Claude Code mods from inside Obsidian, across every coding tool you use: Claude Code, Cursor, Codex, Gemini CLI, and more.
Latest update: Mods. AI Skills Manager now manages Claude Code mods: see every mod you have, turn them on and off, promote one you're building into a finished mod, install mods from GitHub, update them with a diff, and roll back from history. See Mods for details.
If you've written a good skill for one tool and then can't find it again, or you keep hand-copying the same prompt into every project and every agent, this plugin turns your scattered ~/.claude/skills, ~/.cursor/rules, .github/prompts, and similar folders into one searchable, taggable library, without moving your files out of the places those tools actually read from.

AI Skills Manager works directly on the real folders each tool reads. It doesn't import your skills into a separate managed copy: enabling, disabling, tagging, and organizing all act on the files in place.
One library, every tool
Scans the real, tool-specific folders for skills, agents, commands, and rules across 15 coding tools out of the box (see the table below), plus the cross-tool ~/.agents/skills shared convention, at both the global (home directory) level and per-project. Every path is editable in Settings if a tool changes its layout or you use a nonstandard setup.
Enable/disable that actually works Toggling an item physically moves it into (or out of) a sibling disabled folder next to it. It's symlink-aware, so a project-linked item stays correctly linked either way. The tool genuinely stops seeing it, rather than a checkbox that only lives in the plugin's own memory.
Grid or list view Switch the library between cards and a compact list from the toggle next to Sort. The list shows each item's tool, type, scope, sessions, and last used at a glance, color-coded so busy and idle items stand out. Click a column header to sort by name, most used, or recently used. The choice is remembered.

Tags, favorites, and collections Organize items with your own tags, star favorites, and group related skills/agents/commands/rules into collections that span tools and projects.
Vault-native metadata Per-item metadata (tags, favorites, collection membership) is stored as plain frontmatter in small markdown notes inside your vault, not hidden in a JSON blob. That means it syncs via whatever you already use to sync your vault, and it's queryable from Dataview or Bases like any other note.
A detail panel that answers the important questions first Opening an item shows its name and type, the tool (and plugin) it belongs to, and its description. Below that, a row of key numbers: estimated tokens while it's available, tokens loaded when it's invoked, sessions over the last 26 weeks, and when it was last used. A daily usage heatmap spanning the last 26 weeks appears once there's any use to chart. The plugin keeps its own daily usage history for Claude Code and Codex, so it outlasts the tools' own transcript cleanup.
Everything else sits in one properties panel: tags, whether it runs automatically or only when called, who manages it (you, a GitHub source, a plugin, or the tool itself) and what that means for edits and updates, version, any other frontmatter fields, file size, path, and where it's symlinked from. For items installed from GitHub, the panel ends with the source repo and its Check for updates and Restore installed buttons. The History button next to Edit opens the item's saved versions.
Integrity checks Problems that make a tool skip an item or never pick it up are flagged on the item and collected on the Insights Health page: broken or missing frontmatter, a name that doesn't match its folder, a missing or overlong description, a broken symlink, a link to a bundled file that isn't there, or a memory its index never points to.

Agent memories
Memory files an agent writes for itself are listed under Memories & Rules, next to your rules. Claude Code's auto-memory (~/.claude/projects/<project>/memory/) is picked up by default and matched to your project workspaces. Disabling a memory also removes its line from the folder's MEMORY.md index, and re-enabling puts the exact line back, so the agent really stops (or starts) loading it. Other tools' memory folders, global or inside each project, can be set per tool.
Link a global skill into a project Add a global skill to a project workspace and it's symlinked into that project's local tool folder, never copied, so it can't drift out of sync with the source. Remove it and only the link goes away; the original is untouched.
Discover and install from GitHub
Point Discover at a GitHub repo (or a specific subfolder) and it walks it for SKILL.md files, agent/command/rule markdown, and Claude Code mods, showing star counts and previews before you install. Filter the results by type to see only what you're after. Or paste a repo URL (or a GitHub tree URL for one branch/subfolder) directly into "Install from GitHub" and pick which tool, type, and project (or global) it lands in. Self-hosted GitHub works too: paste the repo or tree URL from your company's GitHub the same way.


Check for updates, with a real diff
For anything installed through the plugin, "Check for updates" fetches the source repo and shows exactly what changed before you apply anything. Every changed file gets its own tab, with its status and line counts, so an update that touches references/ or scripts/ shows those diffs too, not just the main manifest. Long unchanged stretches fold away, the arrows in the bottom bar (or J and K) step from one change to the next across every file, and [ and ] switch files. "Restore installed" reverts an item back to the exact commit it was installed at. An optional background interval can check every tracked source on its own and flag what's stale, without ever applying an update for you.
On the All page, Check for updates checks every tracked item at once. An item only counts as stale when its own files changed, not just because something else in the same repo did. Items that do have updates are shown on their own right away (the Updates available source filter), and opening one shows a banner with a Review changes button that opens the same diff.

Version history A version of an item is saved before every edit you make in the detail panel and before every GitHub update or restore, so nothing you change here is a one-way trip. Open History from the detail panel to see each saved version, named after the change that followed it ("Before edit to SKILL.md", "Before GitHub update to 6e802bd") with its date and commit.

Open a version to see what restoring it would change, in the same tabbed diff as an update review, then restore it in one click. Restoring saves the current state first, so a restore can be undone too. Restoring a version of a GitHub-installed item also puts its tracked commit back, so "Check for updates" stays accurate. The last 20 versions of each item are kept, in the plugin's own folder.

Claude Code and Codex plugin awareness Reads Claude Code's installed-plugins registry and Codex's installed plugin cache so skills, agents, and commands bundled inside installed plugins show up in the library too, tagged with the plugin they came from. Codex plugins are read-only here because AI Skills Manager has no supported Codex setting for enabling or disabling an installed plugin. Manage Codex plugins from Codex itself.
Installed bundles are browsable from Extensions → Plugin bundles, with search, tool/group/tag filters, sorting, and a breadcrumb back from a bundle's item list.
Insights A collapsible sidebar section with four pages, each answering one question:
~/.claude/projects and ~/.codex/sessions) for a weekly activity chart and a ranked list of your most-used skills and agents.
On Health and Cleanup, each problem is one row: the item's name, a colored status pill, and a one-line explanation of what's wrong. Expand a row for the details (tool, type, description, location, last edit, and for prune candidates the last invocation and context cost) and the actions that fix it, always in the same order: the fix, Open, then Disregard. Broken symlinks can enable a disabled source or reveal the link in your file manager. An expanded overlap shows both items side by side with their full descriptions, so you can disable, open, or delete either one in place.
Any suggestion can be dismissed with "Disregard" so it stops resurfacing, without touching the item itself. "Show disregarded" on each section lists what you've dismissed and brings any of it back.
MCP servers, read-only
A dedicated page under Extensions lists every MCP server configured across your tools, global and per-project, read straight from each tool's own config file (~/.claude.json, .mcp.json, ~/.codex/config.toml, .vscode/mcp.json, and more). It's visibility only, nothing here can enable, disable, or edit a server, but you can jump straight to its config file to do that by hand.
Built for scanning, not guessing A file-tree preview with per-file size and estimated token count for multi-file skills, a background auto-rescan interval you control, and a reorderable sidebar (by type, plugin, tool, project, or collection) that can hide rows with nothing in them.
A mod is a Claude Code plugin that changes Claude Code itself rather than adding skills: a pane or panel, a band above the prompt, a status line entry, a slash command, or a hook on prompts and tool calls. In files, it's a plugin folder whose hooks/hooks.json names a hooks module. Mods run code in every Claude Code session that loads them, so only install mods you trust.
Open Extensions → Mods to see them all. Each card shows the mod's name and version, and where it loads from. Click one for its description, when Claude Code loads it, and a read-only preview of its files.
Where mods come from
| Source | Where it lives | When it loads |
|---|---|---|
| In development | ~/.claude/dev-mods/<session>/<mod>/ |
Only in the Claude Code session that created it |
| Plugin folder | Any folder named in CLAUDE_CODE_PLUGIN_DIRS, in the env block of ~/.claude/settings.json |
Every new session |
| Installed plugin | Installed with Claude Code's /plugin install |
Every session, while the plugin is enabled |
| Skills folder | A plugin folder inside ~/.claude/skills/ or a project's .claude/skills/ |
Every session |
A mod installed from GitHub through Discover is a plugin folder mod whose card shows its repo (owner/repo) in place of "Plugin folder".
Turn mods on and off
The switch on a mod turns it on or off and does nothing else. For a plugin folder or skills folder mod it sets the mod's entry under enabledPlugins in ~/.claude/settings.json (<name>@inline or <name>@skills-dir), the same setting Claude Code's own /plugin screen uses, so nothing is moved, unlinked, or taken out of CLAUDE_CODE_PLUGIN_DIRS. Installed mods switch through their plugin's entry. A backup of settings.json is made before each change. New sessions pick up the change; a session already running keeps the mod until it restarts.
Make a mod Global or link it into a workspace
The link icon on a mod card opens the same picker as a skill's. Global lists the mod in CLAUDE_CODE_PLUGIN_DIRS, so every session loads it. A mod already in ~/mods is listed from there; one kept anywhere else, such as another repo, is symlinked into ~/mods first and listed through that link. Checking a workspace symlinks the mod into that project's .claude/skills/, so it loads only there. Unchecking removes just that listing or link; the mod's own folder is never moved or deleted. When a mod is Global, a copy in a workspace is skipped, since Claude Code loads one copy per name.
Source repo workspaces A workspace can be a source repo: a folder that keeps skills, agents, commands, or mods at its top level instead of in a tool's project folders. Its page lists everything inside, with its mods under their own heading. Nothing there loads until you link it somewhere. When a symlinked item points into a repo that isn't a workspace yet, its preview offers to add it.
Finish a mod you're building
A mod you build with Claude Code starts in its session's dev folder, so only that session loads it. When it's ready, choose Promote to finished mod from its card menu or detail panel. The mod moves to ~/mods/<name> and is added to CLAUDE_CODE_PLUGIN_DIRS, so every new session loads it.
Card menu
Every mod card has a menu with Reveal in Finder, plus what applies to that mod: Promote to finished mod, Open on GitHub, Check for updates, and Delete. Delete asks first, moves the folder to the Trash (or only removes the link, for a symlinked mod), and takes it out of CLAUDE_CODE_PLUGIN_DIRS if it's listed there. Installed mods have no Delete here: remove those with /plugin uninstall in Claude Code.
Install mods from GitHub
Add a repo in Discover and any mods in it appear alongside its skills and agents. Use the Mods filter to see only those. A mod's preview shows its README followed by the hooks module code, so you can read what it would run before installing. Install copies the exact version you previewed into ~/mods/<name> and adds it to CLAUDE_CODE_PLUGIN_DIRS. Skills bundled inside a mod come with it rather than being listed separately.
Updates and history For a mod installed from GitHub, Check for updates compares it with the repo. If the mod changed, you get the same tabbed diff as a skill update, and nothing changes until you click Update. A version is saved before every update or restore: open History in the mod's detail panel to see them, compare one with what's on disk, and restore it. The last 20 versions are kept.
Insights
Health flags mods that won't load as intended: a CLAUDE_CODE_PLUGIN_DIRS entry pointing at a missing folder or a folder with no plugin manifest, a hooks.json naming a module file that isn't there, and two finished copies of the same mod loading at once (so every hook runs twice). Cleanup lists dev copies of mods you've already finished, and empty session folders under ~/.claude/dev-mods that are more than a day old. Mods don't appear on Usage or Context: Claude Code doesn't record when a mod's hooks run.
| Tool | Skills | Agents | Commands | Rules |
|---|---|---|---|---|
| Claude Code (+ its plugins) | ✓ | ✓ | ✓ | ✓ |
| Cursor | ✓ | ✓ | ✓ | |
| Codex | ✓ | ✓ | ✓ | ✓ |
| OpenCode | ✓ | ✓ | ✓ | |
| Antigravity | ✓ | ✓ | ✓ | ✓ |
| GitHub Copilot | ✓ | ✓ | ✓ | |
| Cline | ✓ | ✓ | ✓ | |
| Trae | ✓ | ✓ | ||
| Windsurf | ✓ | ✓ | ||
| Goose | ✓ | |||
| Hermes | ✓ | |||
| Pi | ✓ | ✓ | ||
| Gemini CLI | ✓ | |||
| Roo Code | ✓ | |||
| Continue | ✓ | ✓ | ||
Shared (~/.agents/skills) |
✓ |
Rules is a mix of directories of many rule files (e.g. Cursor's .cursor/rules/) and single instructions files that are scanned and toggled as one item (e.g. Claude Code's CLAUDE.md, Codex's AGENTS.md, Antigravity's GEMINI.md and AGENTS.md, Pi's AGENTS.md). Codex's AGENTS.md files are persistent instructions, not Claude-style auto-memory files; no Codex memory directory is assumed by default.
Checkmarks reflect the paths scanned by default. Every one of them, plus a few documented-but-unconfirmed guesses for newer tools, can be added, changed, or turned off per tool in Settings.
Support for additional tools and edge cases is welcome. If you use one of these tools and can test a layout, share feedback, or improve its scanner, see CONTRIBUTING.md.

AI Skills Manager is desktop-only: it reads and writes files directly on disk (and shells out to git for the GitHub-powered Discover, install, and update features), which isn't available in Obsidian's mobile sandbox.
Deleting an item, or updating one from GitHub, moves the old copy to your system Trash instead of erasing it. On macOS the plugin asks Finder to do this so Put Back works, which means macOS asks once for permission to let Obsidian control Finder. If you decline, items still go to the Trash, just without Put Back.
AI Skills Manager).Discover, install, and update checks run git in the background. On a managed work computer, two things can get in the way:
HTTPS_PROXY) or your Mac's network settings automatically. If it still fails, find the proxy address with echo $HTTPS_PROXY, git config --get http.proxy, or scutil --proxy in Terminal, or ask IT for "the HTTP proxy host and port for git", and enter it as http://host:port under Settings → Git proxy.git clone <repo URL>) and sign in when asked, so git saves your credentials. Adding the source in the plugin works after that.If you see a certificate error instead, your network inspects secure traffic. Ask IT for its root certificate and point git at it with git config --global http.sslCAInfo <path to certificate>.
npm install
npm run dev # watch build
npm run build # type-check + production build
npm run lint
Contributions are welcome, especially from people who use tools we cannot test locally. See CONTRIBUTING.md for setup instructions, scanner guidance, fixture patterns, and testing expectations.
MIT