The brainstem’s Browse panel lists 138 agents with descriptions, tags, authors, categories, and one-click install. There’s no marketplace backend. There’s no API. There’s a JSON file at the root of a GitHub repo.

https://raw.githubusercontent.com/kody-w/RAR/main/registry.json

That URL returns a single document containing every agent in kody-w/RAR’s library. The brainstem fetches it on boot, parses it, and renders the rows. New agent published? Re-fetch (or wait for the in-page refresh). No deployment, no schema migration, no CDN purge.

The schema is rapp-registry/1.0. It looks like:

{
  "schema": "rapp-registry/1.0",
  "version": "1.0.0",
  "generated_at": "2026-04-18T02:35:25Z",
  "stats": {
    "total_agents": 138,
    "publishers": 7,
    "categories": 19
  },
  "agents": [
    {
      "schema": "rapp-agent/1.0",
      "name": "@aibast-agents-library/account_intelligence",
      "version": "1.0.0",
      "display_name": "Account Intelligence",
      "description": "360-degree account briefings with stakeholder mapping...",
      "author": "AIBAST",
      "tags": ["b2b", "sales", "stakeholder-mapping"],
      "category": "b2b_sales",
      "quality_tier": "community",
      "_file": "agents/@aibast-agents-library/b2b_sales_stacks/account_intelligence_stack/account_intelligence_agent.py",
      "_seed": 4997715477691771520,
      "_sha256": "48010aaae45ac5e099bea276e3b90b10ee9e2e6ab03722f59f423c0d0034cc0c",
      "_has_card": true
    },
    ...
  ]
}

The fields without underscores are part of the agent’s __manifest__ — what the agent itself declares. The _-prefixed fields are derived by the registry build script: file path, content hash, card seed, size. The brainstem only needs _file to install: it fetches https://raw.githubusercontent.com/{owner}/{repo}/{branch}/{_file}, mints the card, adds it to the binder.

That’s the entire installation pipeline. Two HTTP GETs. No package manager.

Browser-side, the brainstem can pull from any number of sources:

state.settings.sources = ['kody-w/RAR'];

In Settings, the user types owner/repo to add another. The brainstem fetches each source’s registry.json (trying main then master), aggregates all the agents into one filterable list, and tags each with its source. Every Browse row carries a “kody-w/RAR” or “your-org/internal-agents” chip so the user knows where it came from.

The registry isn’t a database. It’s a generated artifact. RAR’s CI rebuilds registry.json when agents change (build_registry.py walks the agents folder, computes hashes, mints cards, writes the JSON). Static GitHub Pages serves it. Edge caches keep it fast.

What this means in practice:

  • Forking is trivial. Clone RAR, delete agents you don’t want, add agents you do, push. Your fork has its own registry.json. Anyone can use it.
  • Private libraries are free. Make a private repo with the same layout. The brainstem won’t be able to fetch it (no auth on the unauthenticated raw.githubusercontent.com URLs), but a tiny worker can proxy with a token. Or just use a public repo with obscure naming.
  • There’s no central authority. No one approves what goes into “the marketplace” because there is no marketplace. There’s a default source (RAR) and however many sources users add.

The whole thing fits in a few hundred lines of build script (server-side) and a few hundred more in the brainstem (client-side). It scales as well as GitHub Pages does, which is to say: well past the size of any single project.

If you’re tempted to build a registry server, build a JSON file first.