Your AI tools can now use your Takt library as context

There's a tax you pay on every good AI conversation, and it's paid in copy-paste.

You've read the three articles that matter. You saved the podcast episode where someone finally explained the thing properly. You've got the PDF, the Reddit thread, the half-remembered YouTube talk. Then you open Claude or ChatGPT to actually think about it, and you start pasting. A paragraph here, a link there, a summary you write from memory because finding the original would take longer than approximating it.

The model doesn't know what you know. So you re-type your own brain, badly, at the start of every chat.

What changed

Takt now runs a remote MCP server. Point any MCP-capable AI client at https://mcp.trytakt.com/mcp, sign in once, and that assistant can search your Takt library, pull the relevant notes, and reason over them with citations back to the original source.

Not a summary you pasted. Not an export you uploaded last month. The live library, queried at the moment the question is asked.

This works in Claude Code, Claude.ai, ChatGPT, Cursor, Codex, Grok, and anything else that speaks MCP. One endpoint, every client.

What you can do with it

The obvious thing first: ask questions across everything you've saved.

"Search my Takt library for what I've read about pricing psychology and give me the three ideas that actually conflict with each other."

Then it gets more useful, because the assistant isn't just retrieving — it's working.

Write with your own sources. "Draft the intro to my post on AI note-taking, based on my Takt notes, and cite which note each claim comes from." Your AI pulls an evidence pack of excerpts and writes against it, so the draft is grounded in things you actually read rather than things the model half-remembers from training.

Decide with what you've collected. "I'm choosing between Postgres full-text search and a vector DB. Use my Takt library as context and tell me what I've already read that bears on this." Six months of saved reading, applied to the decision in front of you.

Code against your own reference material. In Cursor or Claude Code, the docs, spec pages, and blog posts you clipped are one tool call away from the file you're editing. No tab-hunting.

Run a weekly review. "List what I saved in Takt this week, group it by theme, and tell me what I'm clearly obsessed with lately." Uncomfortable, occasionally.

What your AI can actually see

Worth being precise about this, because "connect your notes to AI" can mean a lot of things.

Once connected, your client gets seven tools:

• gather_context — a compact, citation-ready evidence pack for a task you describe. This is the one that does the heavy lifting when you say "based on my Takt library."
• search_notes — hybrid search across notes, collections, and tags. Lexical-first, with semantic recall for conceptual queries.
• get_note — the full Markdown of a single note.
• list_notes — recent library metadata, no bodies.
• list_collections — your folders and their note counts.
• ask_takt — Takt's own multi-note synthesis, for when you want Takt to answer rather than the client.
• whoami — which Takt account this session is signed into.

Everything here is read-only. Connected assistants can search your library and read your notes; they cannot edit them, delete them, or save new ones. Your library changes when you change it.

Access runs through Takt's own OAuth. The client receives an MCP access token issued by Takt, never your underlying database credentials, and every query runs under row-level security scoped to your account. Nobody else's notes are reachable, including by you.

Why this matters more than it sounds

Saving things and thinking about things have been two separate activities on two separate days. You collect in one app, patiently, over months. You think in another app, in a burst, with whatever happened to be in your head at the time. The collection is slow and cumulative. The thinking is fast and amnesiac. They almost never meet.

The gap between them is where most of a reading habit goes to die — not in the saving, which people are good at, but in the retrieval, which nobody is good at. An article you read in March is functionally gone by August unless something puts it back in front of you at the exact moment it's relevant.

MCP is what closes that gap. The library stops being an archive you visit and becomes context the model reaches for on its own, mid-thought, without you remembering that the relevant thing exists.

How to connect it

You'll need a Takt account. Everything below takes about a minute.

claude mcp add --transport http takt https://mcp.trytakt.com/mcp

{
  "mcpServers": {
    "takt": {
      "url": "https://mcp.trytakt.com/mcp"
    }
  }
}

codex mcp add takt --url https://mcp.trytakt.com/mcp
codex mcp login takt

Two things that trip people up

The MCP sign-in uses your Takt email and password. If you created your account with Google, you won't have one yet. Use "Forgot password" on the Takt sign-in screen to set one, then authenticate.

If the tools appear but every call fails, you're connected but not signed in. Open your client's MCP settings and run the Takt authentication step again — that browser round-trip is what issues the token.

To check it worked, ask your assistant: "What's in my Takt library?" If it comes back with your own notes, you're done.

What it costs

Library access over MCP is included on every plan, Free included. Searching, reading notes, listing collections, and pulling evidence packs are unmetered.

The one exception is ask_takt, which runs Takt's own synthesis model and draws on your monthly Ask Takt allowance — 10 questions on Free, 100 on Plus, shared with the Takt apps. Everything the connected client does with gather_context and search_notes leaves that allowance alone, because the reasoning happens in your client, not ours.

Connect it from your integrations settings, or create an account and save a few things first. The library gets more useful the moment your AI can reach it.