AgentGem — Distribution

One neutral Gem fans out to code-gen targets, the GitHub registry, and managed deploy backends

Gem manifest + lock the neutral source materialize(gem, target) per-type renderers + compose hook → FileTree Code-gen targets Eve · Flue OpenAI Sandbox AgentCore (harness) claude · codex · agents · hermes Run / deploy eve start (local) Vercel (eve) Cloudflare (flue) src/gem/run.ts Registry publish · resolve merge · install semver + dep graph GitHub repo registry.json (index) items/<scope>/<name>/<ver> atomic commit (Git Data API) Other machines resolve + merge install → workspace or materialize Deploy backends DEPLOY_REGISTRY preview · ready · deploy idempotent publish + undeploy Anthropic Managed Agents skills + environment + agent AWS Bedrock AgentCore CreateHarness (poll status) Deploy record .deploy/<backend>.json ids · urls · harnessId drives Undeploy buttons Gem / record (state) Compose Targets / registry / deploy External backends

Targets share a compose step

  • • materialize() runs per-type renderers, then compose
  • • Eve / Flue / OpenAI Sandbox are code-gen
  • • AgentCore is code-gen + a managed backend
  • • Adding a target ≈ describing how it materializes

Registry is just GitHub

  • • registry.json index + items/<scope>/<name>/<ver>
  • • Atomic publish via the Git Data API
  • • resolve → semver + dependency graph
  • • merge → reconcile manifests, re-resolve one lock

Deploy records drive state

  • • Each deploy writes .deploy/<backend>.json
  • • Records hold ids, urls, harnessId
  • • Undeploy reads the record and tears down
  • • The UI's Undeploy buttons reflect real state