dsh-feed

Plugin distribution markets share a structural inefficiency: every platform scrapes and maintains its own siloed dataset, closed-source and redundant. An agent that needs to query plugin metadata across ecosystems ends up writing multiple scraping scripts, with no single diffable artifact to rely on. dsh-feed deliberately steps back from the UI layer and builds an open data foundation instead, normalizing the GitHub dsh-plugin topic and the npm registry into one plain-text JSON index that downstream tools can consume without permission overhead.

Three equivalent query surfaces are provided: model-side tools (feed_sync, feed_search, feed_stats), a dsh-feed CLI exposing sync, search, stats, and mcp subcommands, and an experimental stdio MCP server that lets any MCP client such as Cursor or Claude Desktop issue natural-language queries over the index. The index file lives inside the session workspace under .dsh/ by default; it is dependency-free, human-readable JSON, and network access is confined to the explicitly triggered feed_sync call so the query path stays fully offline. Search supports mixed English keywords and Chinese bigram matching out of the box.

It is well suited to developers building agent workflows that need a unified plugin search across GitHub and npm, to teams embedding a plugin catalog into IDEs or desktop apps without reinventing the scraping layer, and to MCP power users who want a single, open data source behind their tool stack. The roadmap points toward scheduled hourly syncs, a read-only Web API, and a shared data layer with other dsh ecosystem plugins to further eliminate duplicate fetching.