The brainstem reads any GitHub repo that publishes a registry.json at its root in the rapp-registry/1.0 schema. The default is kody-w/RAR, but Settings → Agent sources lets users add owner/repo entries. Each becomes a tab in the Browse panel.

The 5-minute path to running your own:

1. Fork kody-w/RAR.

gh repo fork kody-w/RAR --clone --remote
cd RAR

2. Throw away what you don’t want.

# Wipe the upstream library, keep the structure.
rm -rf agents/@aibast-agents-library agents/@borg agents/@kody-w

The agents/ directory is organized as agents/@<publisher>/<stack>/<agent>_agent.py. Keep the structure even if you only have one publisher namespace — the registry script expects it.

3. Drop in your own agents.

mkdir -p agents/@your-team/general
cp ~/your-cool-agent.py agents/@your-team/general/

The agent file just needs to follow the single-file contract — __manifest__, a class extending BasicAgent, a perform() method.

4. Rebuild the registry.

python build_registry.py

This walks agents/, parses each manifest, computes hashes, mints cards (saves them to docs/api/v1/cards/), and writes the new registry.json. About a second per agent.

5. Push.

git add -A && git commit -m "internal team agents v1"
git push

That’s it. Your registry is now live at https://raw.githubusercontent.com/your-team/RAR/main/registry.json.

6. Add it as a source in the brainstem.

Settings → Agent sources → type your-team/RAR → Save. Browse now shows your agents alongside (or instead of) the default RAR library.

For private libraries, the path is similar but you have to think about access. raw.githubusercontent.com doesn’t authenticate, so a private repo’s registry.json won’t be reachable directly. Two options:

  • Make it public with an obscure name. Probably fine for most internal-team libraries — there’s no PII in agent code, just descriptions and parameter schemas.
  • Run your own proxy worker that takes a GitHub PAT, fetches from the private repo, returns to the brainstem. This is a ~30-line Cloudflare Worker. Add it as a source like your-proxy.workers.dev/your-team/RAR.

For team-internal stuff specifically, the workflow is: each agent is a .py file in a private repo, the team curates which ones make it in, the registry rebuild runs in CI on push, and team members install agents from the brainstem’s Browse panel just like they would the public library.

The pattern: distribution is a JSON file. Curation is a git repo. Discovery is a brainstem source entry. None of these involves your team running a server.

A team’s “agent shop” is one repo that everyone trusts. Forking that repo is how teams customize. Sourcing different repos is how teams compose. The model scales linearly with the number of repos, which is to say it doesn’t bottleneck — every repo is independent.