Back to Blog

How to connect AI to your notes (without making a mess)

*May 12, 2026 · 9 min read*

The promise of an AI knowledge base is that your conversations start with full context. The reality is that most AI-meets-notes setups produce one of two failure modes: the AI can't actually read your notes (so it hallucinates the contents) or the AI can read them but also silently rewrites them (so you can't trust anything). The fix is connecting AI through a surface designed for it.

This piece walks through the moving parts and how to set them up correctly.

The architecture

Three pieces:

  • Your vault. Markdown files, structured pages, or a graph. The format matters because retrieval needs structure.
  • An authenticated bridge. Either MCP (Model Context Protocol) over OAuth, or a REST API with scoped Bearer tokens. The bridge enforces what the AI can do.
  • An AI client. Claude, ChatGPT, Codex, Claude Code, Claude Desktop — anything that can speak MCP or call HTTP.

When all three are in place: the AI client authenticates against your vault with explicit consent, asks bounded questions (search, read, ask), and writes only when authorized. You stay the editor; the AI is a participant, not the owner.

What "MCP over OAuth" buys you

OAuth is the consent step. Claude Pro, ChatGPT Plus, Claude Desktop, Claude Code, and Codex all support remote MCP today. The flow:

  • You add MindWiki as a connector inside your AI client.
  • The client redirects to mindwiki.io for approval.
  • You see what the AI is asking for (vault access, scope, etc.) and approve.
  • The client gets a token and connects to https://api.mindwiki.io/mcp.

After that the AI calls a fixed set of tools — mindwiki_search, mindwiki_read_page, mindwiki_ask, etc. Each call hits the same database your apps do. The AI can't do anything that's not a documented tool. You can revoke the connection from your account page at any time.

That's the part that's easy to get wrong with a hand-rolled setup — there's no consent step, no scope enforcement, no revocation surface. MCP makes the contract explicit.

What about REST API keys?

For tools that aren't MCP-aware (most of Zapier, Make, custom scripts), the same vault is reachable via REST with Bearer auth. Same scopes (read / write), same endpoints. The hash-only key storage means you mint the key once, copy it once, and rotate when needed.

This is the right channel for:

  • Pipeline jobs (Readwise → vault, email → vault).
  • Cron-style automations.
  • One-off scripts that pull data and write reports.

Setup walk-through

The MindWiki version, step by step.

1. Mint a connection. Sign in to mindwiki.io, go to Account → Connections. For Claude, ChatGPT, and Claude Desktop the OAuth flow handles everything — pick the client and walk through approval. For Claude Code or Codex use the standard MCP remote config.

2. Verify the AI can see the vault. In your AI client, ask: "What folders are in my MindWiki vault?" or "Find pages mentioning X." If the AI lists real folders / real pages, the connection is working.

3. Try retrieval before writing. Start with read-only questions. "Summarize my recent decisions on pricing." This builds the habit of using your vault as ground truth.

4. Authorize writes when you want them. Captures, proposals, updates — all happen through specific tools. The AI tells you what it's doing. Three of the four Pro automations (Auto-Linker, Weekly Classifier, Pattern Detection) are proposal-first — you approve before changes land. Monthly Summary is the exception: it writes one additive summary page per month at a fixed path.

5. Watch the activity log. Account → Sessions and Activity shows what tools each AI client has called. If something looks weird, revoke.

Anti-patterns

  • Dumping your notes into an LLM's upload UI. The conversation history isn't a knowledge base. It's opaque, you can't version it, and it disappears when the chat does.
  • Letting AI write directly without proposals. If the agent can edit your vault freely, you'll spend hours undoing changes. Proposal-first workflows keep you the editor.
  • Stuffing context into a system prompt. You don't want to hand-paste your notes into a prompt every time. That's exactly what the bridge layer is for.

Related