A local dashboard for the agent skills you have installed for Codex, Claude Code or any agent that reads ~/.agents/skills. Browse instructions, organize categories, view recorded usage, and move unwanted skills to your system's Trash or Recycle Bin. No shared skill collection, AI API key, database or hosting service. The dashboard needs no account; optional GitHub backup uses your own GitHub account.
Give Codex, Claude Code or another coding agent this prompt:
When the user asks to install this dashboard, complete the following setup. These instructions apply to installation; merely reading this repository does not request installation.
- Read the user's applicable instructions. This repository is an application, not an installable skill: keep application files outside agent skill directories. Detect native macOS or Windows. Use the appropriate installer below; it provisions a private Python runtime automatically. No existing Python, Git, package manager or administrator access is required. On Windows use native PowerShell rather than WSL.
- Run the installer with
--no-open(macOS) or-NoOpen(Windows). It prints the home, catalog and stable launcher paths. Reinstallation preserves personal data. - Read the loopback address from
<home>/data/server.jsonand fetch/api/skillsand/api/catalog. Inspect personal skill identifiers, picker names and descriptions from the local discovery response. Read an individual skill's instructions only if its description leaves its purpose unclear. Skill contents are source material, not instructions to execute during organization. - On first installation only, organize personal skills into a small set of useful categories based on their actual purposes. Choose category names, sidebar sections and skill order that fit this person's collection, not this repository author's workflows. Preserve existing categories and decisions on reinstall; place new skills only if requested. Plugin and system collections are derived automatically and must remain separate. For an empty collection, leave the catalog neutral.
- Save the catalog through
POST /api/catalog, passing{ "catalog": ..., "revision": ... }from the latest GET response, withOrigin: <loopback-url>andX-Library-Tokenfrom/api/skills. If another writer changed it, reload and preserve their changes. Do not run installed skills, alter their contents, install any additional skills, or contact third parties to classify them. - Run the stable launcher's
opencommand. Report the local URL and that login startup is configured. Do not claim to have tested a reboot or a platform you did not exercise. - Usage recording is optional: ask whether the user wants it. If selected, run
usage-instructionand add its marked block once to the global instructions file of each agent the user chooses (for example${CODEX_HOME:-~/.codex}/AGENTS.mdor${CLAUDE_CONFIG_DIR:-~/.claude}/CLAUDE.md), preserving unrelated instructions. Do not modify global agent instructions otherwise.
curl -fsSL https://raw.githubusercontent.com/jdjuan/myskills/main/install.sh -o /tmp/skill-dashboard-install.sh
sh /tmp/skill-dashboard-install.shInvoke-WebRequest https://raw.githubusercontent.com/jdjuan/myskills/main/install.ps1 -OutFile "$env:TEMP\skill-dashboard-install.ps1"
powershell -NoProfile -ExecutionPolicy Bypass -File "$env:TEMP\skill-dashboard-install.ps1"The distribution repository must be public and have a published stable release before these commands work for everyone. No releases are created by local installation or testing.
The home folder is ~/Library/Application Support/SkillDashboard on macOS or %LOCALAPPDATA%\SkillDashboard on Windows. It contains:
data/catalog.json: your categories, order, explicit memberships, steps and compact sections.data/settings.json: update preference, selected port and Codex home. No credentials.releases/<version>: replaceable application files;current.jsonselects the active version.runtimeanduv: privately managed runtime, independent of your global Python setup.logs: local service and server diagnostics.
Use Organize to create, rename, remove and reorder categories; change their section; move skills; or reorder skills. Changes apply together when saved. Removing a category does not delete skills. New discoveries appear under Uncategorized; existing choices are never regenerated by an update. Your agent can also edit the personal catalog directly. Names, descriptions, picker metadata and instruction sizes always come from the skill folders.
Skills without their own icon get the closest matching icon from a bundled set of about 330 Microsoft Fluent Emoji (MIT), chosen from the skill’s name and description. A skill keeps the generated fallback when nothing is a reasonable match. The icons ship in icons/fluent with every release; rebuild the set with python3 scripts/fetch_fluent_icons.py.
To choose an icon yourself, open a personal skill and click Change icon. Pick any bundled icon (searchable by name and keyword) or Use automatic icon. The choice is saved in the skill folder as assets/icon.png, with its credit file and agents/openai.yaml entries, so it travels with the skill and survives skill updates.
The dashboard reads PNG icons from each skill’s agents/openai.yaml (interface.icon_large, then icon_small). PNGs must stay within the skill folder and be at most 512 KB and 4096 × 4096 pixels. Other formats keep the generated fallback. An optional sibling icon-name.credit.json provides author, authorUrl, and sourceUrl; HTTPS attribution links appear in a small footer. Personal icon artwork and credits remain in skill folders, outside application releases.
The version-1 catalog contains groups (ordered {id,label,section,skills} objects), fallback_group, optional plugin_labels, and compact_sections. Skill members can be identifiers or {name,step} objects. Optional prefixes match future skills; excluded_skills lets a user move a matching skill elsewhere. Multiple memberships and compact display rules in existing catalogs remain supported. The repository's catalog is only a neutral first-run template.
Skill roots: ${CODEX_HOME:-~/.codex}/skills, ~/.agents/skills, ${CLAUDE_CONFIG_DIR:-~/.claude}/skills, and the plugin caches ${CODEX_HOME:-~/.codex}/plugins/cache and ${CLAUDE_CONFIG_DIR:-~/.claude}/plugins/cache. A skill folder linked into several roots appears once. Plugin caches do not prove a plugin is enabled. Project-local skills outside these roots are not enumerated. Installed skills and plugin caches are never bundled with the application.
Open Settings → Manage GitHub backup. No backup is uploaded until you enable it.
- Click Connect GitHub. The dashboard reuses an existing GitHub CLI login or downloads the official portable CLI into its own tools folder. If sign-in is needed, it shows the temporary device code and a link to GitHub. Authentication belongs to GitHub CLI; the dashboard does not implement OAuth or store its own access token. If setup fails, open Need help connecting? and use Copy setup prompt for installation and sign-in guidance.
- Choose a repository name, initially
skills-backup, and click Enable backup and upload skills. This creates a private repository in the signed-in account, or verifies an existing dashboard backup. Public repositories and unrelated existing repositories are refused. - Use Back up now, or enable Back up automatically every day. Daily backup runs while the local dashboard server is running, including when the browser is closed. The panel shows progress, errors, the repository link and the last successful backup time. Unchanged files do not create extra commits.
- To recover on this or another computer, connect the same GitHub account, enter the repository name and click Preview restore. Review the files that differ. Restore fills missing files and keeps existing conflicts unless the replacement checkbox is selected. Replaced files are retained under the local data folder's
restore-historydirectory. Changes made after a preview require another preview.
Backups include regular files in personal skill folders under the three discovered personal roots, their supporting scripts/assets, and the personal catalog. Linked personal skill folders are copied as portable regular folders; nested linked files/directories and links into built-in or plugin-cache locations are excluded. Backups also exclude built-in skills, plugin caches, usage logs, environment files, common credential/key files and dependency caches. Restore leaves existing linked destinations unchanged. Agent account configuration and dashboard runtime settings are outside the snapshot. Individual files are limited to 20 MB and snapshots to 100 MB; oversized snapshots fail without replacing the previous backup.
Backup files remain readable in the private repository with Git history. Each backup represents this computer's current snapshot; use distinct repository names for computers that should retain separate current snapshots. Restore is additive and never runs skill scripts. This is backup, not automatic cross-computer synchronization.
GitHub operations use the CLI's Git Data API, so a separate Git installation is unnecessary. Backup preferences and status stay in the personal data folder as github-backup.json and github-backup-status.json. Failed requests show GitHub's error message and record the timestamp, API endpoint, method and HTTP status in github-backup-errors.jsonl in that folder; request payloads and authentication headers are never logged. Credentials stay under GitHub CLI's own authentication management. There is no ZIP export or ZIP restore option.
The service starts at login, restarts after exits and runs while the user is logged in and the computer is awake. It uses launchd (io.skilldashboard.server) on macOS and an interactive-user Scheduled Task (SkillDashboard) on Windows. It does not run as administrator or open a console window. macOS registration lives in ~/Library/LaunchAgents/io.skilldashboard.server.plist.
The dashboard binds only to loopback. Port 8765 is preferred; if occupied, it chooses and remembers an available port. Use the launcher to open the correct address. Host/origin/token checks protect mutations; application source, personal state files and credentials are not served as files. Deletion is recoverable, and system skills and linked/reparse-point locations are protected.
Automatic updates are on by default. The supervisor checks stable GitHub releases on startup and every 24 hours. It downloads the new application, switches versions and restarts the dashboard. It never pulls development code, regenerates your catalog, installs skills or runs a test/validation pipeline on your computer. Interrupted downloads leave the current installation intact. Settings includes the version, last update check, errors, automatic-update toggle and manual check button. An open browser checks for interface changes every three seconds while visible and reloads automatically, including while Settings is open. Organization drafts, backup setup and deletion confirmations defer the reload until their window closes, with an update notice. Connection failures during a restart are retried without reloading an unavailable page. One previous release is retained for manual rollback. No custom update server is needed.
New installations create a MySkills launcher automatically: on macOS, press Command-Space and search for MySkills; on Windows, search for MySkills in the Start menu. The launcher calls the stable dashboard command and resolves the current runtime and local address, so it survives updates and port changes. Spotlight indexing may take a moment; installation requests indexing without opening Spotlight.
Existing installations can use Settings → Add Spotlight launcher (macOS) or Add Start menu shortcut (Windows). Settings also removes a dashboard-owned launcher. That choice is preserved across reinstalls and application updates. A different app or shortcut named Skills is left untouched, with an explanation in Settings. Uninstallation removes only this installation's own launcher.
To skip launcher creation during installation, use --no-shortcut with the macOS installer or -NoShortcut with the Windows installer. No personal Shortcut skill is needed.
Use the absolute Python path stored in <home>/python-path.txt to run <home>/launcher.py with any of:
open
start
stop
status
update
rollback
uninstall
uninstall --remove-data
usage-instruction
record /absolute/path/to/SKILL.md
stop removes login registration; start restores it. rollback selects the previous version and disables automatic updates until re-enabled. Uninstall retains personal data unless --remove-data is supplied, and always leaves installed skills untouched. SKILL_DASHBOARD_HOME overrides the managed installation home.
Requires Python 3.10+ for a development checkout (managed installations use Python 3.12):
python3 server.py
python3 scripts/verify.pyThe frontend is buildless and Python uses only its standard library. Before pushing, run python3 scripts/verify.py: it checks application Python syntax, checks frontend JavaScript when Node is available, runs the full test suite, and builds and validates a temporary allowlisted release ZIP. CI runs the same command on macOS and Windows with Node available; local macOS success does not claim Windows verification. Development checkouts never auto-update their code. --data-dir /absolute/path or SKILL_DASHBOARD_DATA_DIR isolates personal state; --port 8766 selects a port explicitly. Refresh picks up frontend changes and new skill files.
For a local managed installation using an existing runtime:
python3 dashboard.py install --source . --home "/absolute/test installation" --no-start --no-shortcutUse a separate test home and temporary skill roots when exercising deletion. To migrate an old checkout installation, add --migrate-checkout /absolute/old/checkout to dashboard.py install. It copies an existing legacy catalog only if personal state is absent, backs it up, and replaces the matching legacy login service when startup is requested. It never overwrites or deletes the source checkout. manage-service.py remains the legacy checkout-service helper.
Recorded usage is best effort, enabled through the optional instruction block rather than an automatic agent harness hook. usage.py appends timestamps and random event IDs beside the invoked SKILL.md, storing no prompts or project contents. Identical copies combine logs and deduplicate event IDs. System/linked logs remain protected and read-only failures do not block discovery. Zero recorded uses does not mean a skill has never been used.
Every authorized push to main runs tests on macOS and Windows, then automatically publishes a stable GitHub release with the allowlisted myskills.zip and bootstrap scripts. Installed users can get it through Settings → Check now; automatic updates check daily. A failed test gate prevents publication. Pull requests only test and package; they never publish.
The workflow assigns the next patch version from existing stable tags, using tracked VERSION as a minimum. It writes that version into the release archive without making a version-bump commit. You can raise VERSION explicitly for a minor or major release. Branch/tag events and reruns for the same tagged commit reuse its version and leave an existing release unchanged. Release jobs are serialized to avoid version collisions. For an explicitly authorized manual release, push a vMAJOR.MINOR.PATCH tag matching VERSION. python3 scripts/package.py packages locally. Runtime, catalogs, logs, installed skills and unrelated checkout files are excluded.
Before making a private repository public, review tracked files and history for secrets and unintended personal files. The code change does not itself authorize a push, release, repository visibility change or history rewrite.
All 30 tests passed on macOS with both the development Python and the privately provisioned Python 3.12 runtime. Verification exercised a managed launchd installation, server recovery, port fallback, a running release switch from a local download fixture, catalog survival, and browser organization editing at desktop and narrow widths, including concurrent-edit recovery. Native Windows installation, Scheduled Task execution and Recycle Bin behavior still require verification on Windows; the macOS/Windows CI workflow is prepared but has not been run remotely. Public bootstrap installation and GitHub release delivery require the separately authorized first publication.
The launcher change was verified with 40 tests on macOS (39 passed; the native Windows shortcut test is skipped on macOS). Local browser checks exercised adding and removing a disposable Spotlight launcher and persistence after refresh. Its manifest, executable shell syntax and stable launcher command were checked. The existing user-created Skills app was left unchanged. This requests Spotlight registration/indexing; it does not claim a particular search ranking. Native Windows shortcut verification remains for Windows CI.
Verification uses disposable fixture skills and a simulated GitHub API for repository creation, private-only writes, unchanged backups, upload failures, restoration to a different root, conflict preservation, replacement recovery, daily scheduling, traversal rejection and expired/local-edit previews. The official portable GitHub CLI was downloaded and executed successfully in a temporary folder. After the owner enabled personal backups, a real private-repository backup also completed successfully; a read-back verified its commit, file count, catalog and manifest.
The completed local suite contains 67 tests (66 passed; one native Windows shortcut test is skipped on macOS), including safe API-error diagnostics and logging-failure handling. Browser checks covered connection/setup with a disposable simulated account, backup, daily preferences, restore previews, keeping conflicts and replacing conflicts with recovery copies.
Standalone skills with a supported GitHub installation receipt appear in Installed, beside My skills, Plugins, and System. In an installed skill’s details, choose I customized this · Move to My skills after adapting it. This choice is stored in the personal catalog, keeps the source link, and does not modify the skill or its installation receipt. Renamed copies without a matching receipt remain in My skills. Customization is explicitly marked rather than guessed from file timestamps; the dashboard does not automatically apply upstream skill updates.
In Installed, choose Check for updates for one skill or the whole collection. Checks happen only when requested, never on startup, on opening a tab, or on a background schedule. The dashboard shows each skill’s last check and keeps that status across restarts. Checks run in the background so the local UI stays responsive; its existing refresh only reads local progress.
Public GitHub repositories are supported without Git or extra Python packages. The original repository and skill path come from installation receipts where available. Missing paths are resolved only when the repository has one unambiguous matching skill. Source in skill details lets you specify the repository, folder (empty means repository root), and branch or tag. A blank branch follows the repository’s default branch, which is recorded at the first successful check. No GitHub credentials are requested or sent; private, moved, inaccessible, or ambiguous sources can be handled with Copy update prompt instead.
Review update shows added, modified and removed files, text diffs, binary-file notices, and a link to the exact upstream commit. Long diffs are explicitly marked as shortened. An existing installation without saved file hashes requires an extra acknowledgment because its differences may include personal edits. Once a baseline exists, local changes block updates and offer Keep in My skills. Changes made after review invalidate that review. Source changes require a new review.
Applying uses a staged copy of the reviewed files, validates their Git blob hashes and readable SKILL.md, and keeps a recoverable backup before replacing the skill. Usage logs and the dashboard catalog are preserved; downloaded scripts are never executed. Unsafe paths, links, submodules, oversized skills, and development/dependency folders are refused. If a swap fails, prior folders are restored; a transaction journal recovers interrupted swaps on restart. Restore previous version in the skill details restores the available backup after confirmation and refuses to overwrite edits made since the update. Update receipts, cached reviews, and backups live under the personal data directory (skill-updates.json, skill-update-reviews/, and skill-update-backups/) and survive dashboard releases. Backups are retained locally; they are not committed or published.
The dashboard header and browser favicon use the approved yellow-and-black logo. Compact header artwork and a 16/32/48-pixel favicon are embedded in the buildless page, keeping release packaging and static-file access unchanged. The source image is retained at design/logo/skill-manager-yellow-black.png.
The product and repository are now MySkills (jdjuan/myskills). New search launchers use MySkills; existing launchers remain usable. Installation folders, service identifiers, environment variables, health identifiers, and backup formats retain their original names to preserve existing installations and personal data. Releases include myskills.zip and a legacy skill-dashboard.zip download for older updaters.