Your First 10 Minutes

By the end of this tutorial, you’ll have an agent that knows things, learns from you, and remembers everything between sessions.

Step 1: Create your agent

Open your terminal and run:

npm create soleri my-agent

The wizard will ask you a few things — pick a name, describe what your agent does, choose some knowledge areas. Don’t overthink it, you can change everything later.

Once it’s done, build it:

cd my-agent
npm install
npm run build

Step 2: Connect to Claude Code

Add your agent to .mcp.json in your project root:

{
  "mcpServers": {
    "my-agent": {
      "command": "node",
      "args": ["./my-agent/dist/index.js"]
    }
  }
}

Restart Claude Code. Your agent is now running.

Step 3: Ask it something

Your agent comes with starter knowledge out of the box. Try asking it something:

You: “What do you know about error handling?”

Agent: Found 3 relevant entries. Here are patterns for error handling: Use error boundaries at the route level to prevent full-page crashes…

It already has useful things to say — you didn’t have to teach it anything yet. This is the starter knowledge that ships with every agent.

Step 4: Teach it something new

Now teach your agent something specific to your project. Say you just decided that all API errors should return a consistent format:

You: “Capture this pattern: all API errors must return { error: string, code: number, details?: object }. This keeps our frontend error handling simple and predictable.”

Agent: Captured pattern: “Consistent API Error Format”

That’s it. Your agent now knows this rule. It’s stored in the vault — not in your head, not in a doc somewhere.

Step 5: Find it again

Let’s make sure it stuck:

You: “Search for API error patterns”

Agent: Found: “Consistent API Error Format” — all API errors must return { error: string, code: number, details?: object }…

Your pattern shows up right away. The more patterns you capture, the smarter searches get — the agent learns which ones matter most based on how often you use them.

Step 6: Close and reopen

Here’s where it clicks. Close Claude Code completely. Open it again. Ask the same question:

You: “What do we know about API errors?”

Agent: Found: “Consistent API Error Format”…

It remembered. Not because it has a conversation history — because the knowledge lives in the vault, permanently. Next week, next month, it’ll still know this.

How it works under the hood

Your agent is an MCP tool server — it exposes tools that Claude Code can call. When you said “capture this pattern,” Claude Code called the agent’s capture_knowledge tool. When you searched, it called search_intelligent.

The agent doesn’t proactively surface knowledge on its own. Instead, Claude Code decides when to call the agent’s search tools based on your conversation. When you ask about API errors, Claude Code recognizes this is relevant to your knowledge base and calls the search tool. The agent returns ranked results, and Claude Code uses them in its response.

This is why the vault is powerful — it’s not a passive document. It’s a searchable, ranked knowledge store that Claude Code consults whenever your conversation touches a relevant topic.

What just happened

In 10 minutes, you:

• Created an agent that already knows useful things
• Taught it something specific to your project
• Searched and found it instantly
• Closed the session and it still remembered

This is the foundation. Now that you’ve seen the basics, learn the workflow that ties it all together.

---

Next: The Development Workflow — the five-step rhythm for working with your agent: Search → Plan → Work → Capture → Complete. Then dive into Building a Knowledge Base to learn what to capture and how to organize it.