Tutorial · Members only Agents
Build your own AI assistant with Claude Code, step by step
The complete, tested code for a minimal assistant: Node orchestrator, one Claude Code chat per project, Obsidian memory, custom MCP and approval tickets.
Tested on October 17, 2026 in Claude Code 2.1.283 · Node.js 22.23.3 · @modelcontextprotocol/server 2.3.1 · @modelcontextprotocol/client 2.3.1 · zod 4.6.7 · Windows Server 2022. We ran all the code exactly as published: MCP server, orchestrator, queue, and both approved and denied tickets. We haven't tested it on Linux or macOS. How we test
With four Node files (about 330 lines in total) you can have an assistant that takes on tasks, runs them with one Claude Code chat per project that picks up where it left off, saves what it learns in Markdown notes you can open in Obsidian, and asks for permission before anything sensitive. It's the minimal version of Jacob, the assistant we use every day, and we ran the whole thing on October 17, 2026, before publishing it.
Below you'll find the complete code, how to test each piece on its own, what you should see on the first run, how to schedule it, and the errors and security precautions worth knowing before you leave it running unattended.
What the full version includes
- The complete, tested code: orchestrator, custom MCP server, task queue and approvals (about 330 lines of Node)
- The folder structure and the memory: project card, status note and each project's CLAUDE.md
- How to check each piece on its own and what you should see on the first run
- How to schedule it so it works on its own
- Common mistakes and the security checks to do before leaving it alone
The full version is for members
Articles with our own method, the real configuration and the full code are for readers who support the blog with a membership: €4.99 every 3 months. Everything else stays open to everyone, and so does the AI news board.
You are signed in, but your account has no active membership.
Already a member? Sign in
