Getting started
AgentGem runs as a small local server. A browser can't read your agent's config directory (it's sandboxed), so AgentGem serves the introspection and Gem-building operations from your machine — over both HTTP and MCP — while keeping secrets on-device.
Prerequisites
- Node.js (LTS) and pnpm.
- A supported coding agent — Claude (
~/.claude), Codex, or Hermes — with skills, MCP servers, and instructions. You can point AgentGem at a different directory for testing.
Install and run
pnpm install
pnpm build # AgentBack uses legacy decorators — build with tsc, then run dist/
pnpm start # → node dist/index.js
The server starts on 127.0.0.1 and prints its URL. Three things come up from the same
contract:
| Path | What it is |
|---|---|
/ |
The Gem Builder web UI |
/explorer |
Swagger UI for the REST API (from the OpenAPI document) |
/mcp |
The MCP endpoint, so your local agent can call the same tools |
During development, pnpm dev builds and starts in one step.
Build your first Gem
- Open
/. AgentGem callsGET /api/inventoryand renders your config: skills (with descriptions), MCP servers (with transport), and an Include instructions toggle (e.g.CLAUDE.md). - Tick the artifacts you want to bundle and give the Gem a name. On each change the page
calls
POST /api/gemand shows the live, pretty-printedgem.json— secrets already shown as<redacted>. - Download
gem.json(or copy it). That archive is the neutral source every target and the registry consume.
Targeting a different config directory
Both the UI and the API accept a directory override for testing or non-default homes:
curl 'http://127.0.0.1:<port>/api/inventory?dir=/path/to/fake/.claude'
Calling it from your agent
Because every operation is also an MCP tool, your local agent can call inventory and
gem directly over /mcp — no browser required. This is the agent-native path: the same
contract the web page uses is available to the agent that's building the Gem.
Next: read Concepts to understand what's inside a Gem, or jump to Targets & deploy.