Import and publish
- Open Skills in the dashboard as an administrator. Installed opens by default.
- Choose Import manually to select a local folder containing
SKILL.mdor enter a public GitHub directory or skills.sh URL. Open Discover to browse the first 20 entries from the skills.sh all-time leaderboard, search the catalog, or select a skill to import. - Select an installed Skill by name to open its details. The file browser starts with
SKILL.mdand previews Markdown, UTF-8 text, PNG, and JPEG files up to 512 KiB. Select Enabled in the detail page and choose the publication scope. Changes are saved automatically. Global endpoint publishes to/mcpif global routing is enabled. Selected groups publish to their group endpoints and apply the existing route authentication and group visibility. - Connect a compatible client to the published endpoint. It discovers Skills via
skills/list, fetches their manifests viaskills/get, and reads individual files viaresources/read.
HTTPS_PROXY or HTTP_PROXY and honor NO_PROXY. With neither proxy variable set, they connect directly. You can also force a direct connection or enter a custom HTTP(S) proxy URL. Discovery continues to use the server’s default network configuration.
Discovery uses the public skills.sh catalog endpoints and supports Skills from public GitHub repositories. A skills.sh import finds the Skill in the repository’s default branch and pins the import to that commit. If several directories match, enter a specific GitHub directory URL such as https://github.com/owner/repo/tree/main/skill-name. You can also use a commit SHA or URL-encode the slashes in a branch or tag name.
The directory name must match the name in SKILL.md and be unique in the catalog. Import supports up to 512 regular files and 16 MiB of file content per Skill. GitHub imports resolve the source to a commit at import time. Turn off Enabled or delete the Skill to take it off the published endpoints.
For the TypeScript SDK, create the client with versionNegotiation: { mode: "auto" } so it discovers the 2026-07-28 endpoint.
Storage and compatibility
Each Skill is stored underdata/skills/<skill-name>/ relative to the MCPHub working directory. The directory contains the original SKILL.md and attachments at their original relative paths, plus the reserved .mcphub.json catalog metadata file. Include data/skills/ in backups. In Docker, mount /app/data to persist these directories. The catalog supports either settings backend in single-instance self-hosted deployments.
On first catalog access, MCPHub migrates preview JSON snapshots from data/skills/ and <settings-file>.skills/, preserving IDs and publication settings. If multiple snapshots share a name, the existing directory or first migrated snapshot keeps the name. MCPHub retains the others as .conflict files in their original directory for manual recovery.
Both the manifest and its SHA-256 digests describe the original UTF-8 or binary file bytes. If a Skill changes between fetching a manifest and reading a file, clients can refresh the manifest and verify the new file. Existing tools, prompts, and resources remain available through their usual methods.