The install decision for DeepSeek Harness plugins. Not another list.
There are more than twenty directories of dsh plugins. Every one of them answers what exists, by scraping the same GitHub topic and printing a count. The counts range from 20 to 6,685 for the same universe — which tells you the number is a filter policy, not a fact.
This site answers the next question. For each of 6,290 plugins:
| Row | What it says | Where it comes from |
|---|---|---|
| Proof | the file that proves it installs, as path#key, as a link you open |
evidence in the registry |
| Install | the command, and whether it runs code on your machine at install time | npm / repo / path |
| Pulse | days since last push, and how many on the same shelf pushed sooner | GitHub pushedAt |
| Also | how crowded the shelf is, freshest first | the registry's 17 tags |
| Not | what this page is not evidence of | stated, every time |
Nothing here is scored, graded, ranked, or estimated. Every number is a fact in the registry or arithmetic on one. A directory that invents a number is asking to be believed, and being believed is not what this is for.
memory holds 451 plugins: a context-insight panel, a shell bridge, and a
session-deleter all land there. The site says "451 plugins carry the memory
tag" and never "451 plugins do this job", because the tag is what the data
supports and the stronger sentence would be a lie wearing a fact's clothes.
The shelf is still worth showing. Seeing the crowd before you pick from it is more than a card in a grid of 6,290 will ever tell you.
dshworks/awesome-dsh-plugins
— open data, MIT, 98.8% of the dsh-plugin topic decided, and the only
registry in this ecosystem where every verified row names the file it was
verified from. This site is a reader of that registry, not a second source of
truth: npm run data fetches it and derives, it never edits.
- Next 16 App Router on Vite (
vinext), deployed to Cloudflare Workers. Every HTML route is server-rendered, so all 6,290 plugin URLs are real, crawlable pages — the single-page galleries have exactly one. - The Worker owns paths, not the hostname.
dsh.works/,/p/*,/tag/*,/api/*,/assets/*,/_data/*,/llms.txt,/sitemap.xml. Everything else falls through to GitHub Pages exactly as before, so/awesome-dsh-plugins/,/awesome-dsh-themes/,/dsh-meter/and/dsh-crew/are untouched. It is not timidity:dshworks.github.io/*301-redirects todsh.works/*, so a Worker holding every path could not proxy those project sites without looping into itself. - The registry ships as static assets, not in the bundle. It is ~4 MB
against a 3 MB script limit. The Worker reads it through the
ASSETSbinding: an internal lookup, no egress, and no second store that can disagree with the code that rendered the page. Measured before designing. - No database, no auth, no KV. There is no state. Adding one would be adding something to keep in sync with git for data that is already static.
- The search index loads on touch, not on paint. 907 KB is real; most visitors never search, and everything above the box is server-rendered and readable with JavaScript off.
/llms.txt— the map, in one fetch/api/plugins— counts, tags, endpoints/api/plugins/{slug}— one plugin, including its proof and install command/api/tags/{tag}— a whole shelf, freshest first/sitemap.xml— every page
MIT, no key, CORS open. The audience for a plugin directory is people running agents; making them parse HTML would be a strange way to serve them.
npm install
npm run data # fetch the registries and derive
DATA_SOURCE=local npm run data # ...or read the sibling checkouts
npx vinext build
npx wrangler dev # workerd, with the real ASSETS bindingwrangler dev is the honest local test — it runs the same runtime as
production, with the same binding. vinext dev runs on Node and stubs the
binding out.
Not official — dsh is DeepSeek's, at deepseek-ai/deepseek-harness. Not a security review: an install path was checked, not a codebase, and git installs run code on your machine at install time outside any sandbox. Not a quality ranking. Nothing here is sold, sponsored, or promoted.
MIT.