Per-course starting environments for the Manufacturing track of The AI Agent Factory.
Each folder here is the base for one crash course: rules files, MCP wiring, and an env template. On this track you direct, and your coding agent builds the project on top of the base from prompts you paste. Each course's Quick Win walks you through it, and your agent does the setup itself.
| Course | Folder | Download |
|---|---|---|
| Connector-Native Apps | connector-native-apps/ |
connector-native-apps-base.zip |
| Plugins for AI Agents | plugins-crash-course/ |
plugins-crash-course-base.zip |
| RAG on Postgres (pgvector) | postgres-ai/ |
postgres-ai-base.zip |
| Digital FTE | digital-fte/ |
digital-fte-base.zip |
| AI Agent Nervous System | ai-agent-nervous-system/ |
ai-agent-nervous-system-base.zip |
| Workforce with Paperclip | paperclip-workforce/ |
paperclip-workforce-base.zip |
| Eval-Driven Development | eval-driven-development/ |
eval-driven-development-base.zip |
| Owner Delegation with Identic AI | identic-ai/ |
identic-ai-base.zip |
| AI Identity (Sign-In & Agent Access) | ai-identity/ |
ai-identity-base.zip |
The bases share one spine: Neon and Context7 over MCP, database work through Neon MCP only (dev-plane), audit in the same transaction, and skills the agent installs (skill-creator, mcp-builder, plus whatever a course needs). They differ in that skill set, which is why each course gets its own folder. The Paperclip base is the exception: Paperclip ships its own embedded Postgres and is driven through its CLI and REST API, so that base has no Neon, no Context7, and no .mcp.json; the skills it installs are Paperclip's own operator skills (paperclip-create-agent, diagnose-why-work-stopped). The Eval-Driven Development base adds one more spine member: a local, keyless phoenix MCP (@arizeai/phoenix-mcp against http://localhost:6006) alongside Neon and Context7, dormant until the course launches Phoenix. The Identic AI base keeps the Neon + Context7 spine for its governance ledger and verify-before-you-code habit, and adds the official paperclip MCP (@paperclipai/mcp-server, against a local sandbox) as a third server, so an OpenClaw delegate can read and resolve approvals without a hand-written client. The Plugins for AI Agents base is the other exception: it teaches you to build a coding-agent plugin, not a data agent, so it carries no Neon, no Context7, and no .mcp.json spine. Instead it ships a reference/ proven build of the plugin you make (with its own verify.sh green-gate and a runnable MCP server to wire), and the agent installs the official Plugin Structure skill rather than the data spine's skills. The AI Identity base teaches you to build an identity service rather than a data agent, and like every base it is lean — rules, a library of specs/ with adversarial acceptance criteria, the reader prompts, and one shipped skill (agent-identity-issuer, the gap the official skills leave: OIDC/OAuth issuance and agent identity). It carries the Neon + Context7 spine over MCP and adds a course-specific better-auth MCP (for the moving 1.7 / CIMD surface) in .mcp.json. It pre-builds no app: specs/00-set-up-the-base directs the agent to scaffold Next.js + shadcn, install the official Better Auth skills, and pin the Better Auth 1.7 line (where CIMD lives). One twist on the spine: here Neon is the app's own store (@neondatabase/serverless, no native build) and migrations run through Better Auth's CLI, so the Neon MCP is for provisioning and inspection rather than the dev-plane migration pattern; the OAuth/OIDC issuer is the thing you manufacture.
Open a course folder in your coding agent (Claude Code or OpenCode). CLAUDE.md loads AGENTS.md, which carries the standing rules and tells the agent how to install the course's skills and confirm the MCP servers. Then follow the course.
The worked-examples/ folder holds the finished build for a course, for after you have done it yourself. These are full reference solutions (the complete code, verified end to end), not bare bases, so they are deliberately kept out of the release zips. Treat them as the answer key: read them when you are stuck or want to compare, not as your starting point.
| Course | Reference solution |
|---|---|
| Connector-Native Apps | worked-examples/connector-native-apps/ |
| Digital FTE | worked-examples/digital-fte/ |
| Owner Delegation with Identic AI | worked-examples/identic-ai/ |
| Give Your AI Agent a Nervous System | worked-examples/ai-agent-nervous-system/ |
| AI Identity (Sign-In & Agent Access) | worked-examples/ai-identity/ |
Add a folder with the base files (copy an existing one, adjust AGENTS.md's skill list and .mcp.json if the course needs another server). On the next v* tag, CI builds <folder>-base.zip and attaches it to the Release.
For how a base takes a model credential (free-by-default Gemini, one provider-named key, the agent proves it at setup), follow MODEL-PROVIDER-STANDARD.md. postgres-ai is the reference implementation.
Apache-2.0.