←

Use Google Search & YouTube history as agent context in Cursor via Fabric MCP

Fabric3 likes · 1 restack

When I was planning a summer weekend, GPT suggested Ibiza in August.

Only after I wrote a long monologue about weather between 20 and 25 degrees, an interest in history, and a dislike of crowds did the ideas get slightly better.

I finally booked Capri in late September.

The activities suggested by different AI apps did not feel like me because they do not actually know my interests.

I have not even asked them about my upcoming trip to India.

Most AI apps work with thin personal context, so they guess.

So how do we build AI products that pick up evolving interests across platforms, instead of starting from zero?

TL;DR

  • Plug in context with Fabric MCP: with your consent, Fabric MCP lets an agent read high signal slices of your Google Search & YouTube history as context

  • Cold start solved: the agent starts with a real picture of your interests, purchases, and upcoming plans before you type a word

  • Works anywhere MCP works: I am showing Cursor but any MCP client can use the same context

What is Fabric MCP

Fabric MCP is a server that connects providers like Google, YouTube, Instagram, and Gemini to your MCP compatible agent, with your permission. The agent can then use Fabric to request user context across connected data sources.

Setup takes minutes. Connect your data in Fabric. Add Fabric MCP to your MCP client. Prompt in your MCP client.

Connect your data in seconds

I had previously connected my Google data with Fabric, so the consent flow was a couple of clicks.

Add Fabric MCP to your client

Cursor Demo capturing my interests

I asked the agent to create a markdown file of my interests.

Can you generate me a file called interests.md which summarises my interests from Jul-Sep 2025 using Search and Youtube data:

- Demographics: age, gender, ethnicity, country etc.
- Interests: professional, tech, fashion, culture, dining etc.

It did a great job. The summary reflected my interests accurately. It works because the agent isn’t hallucinating my personality. It’s reading real signals from my online behaviour.

The important bit is what this unlocks. Any agent can start with a nuanced understanding of the user, without waiting for thousands of interactions.

What it found for me (real examples)

Here are some inferred preferences pulled from interests.md. The video goes deeper. This came from a simple prompt, and the agent can drill into any detail by querying the user’s interactions via Fabric MCP.

Q3 2025

AI agents, tooling and infra: A2A, Google ADK, MCP, n8n, Vapi, zep, Mem0, Fabric, Cursor, OpenAI

Data portability and tech policy: EU Digital Markets Act (incl. Art. 6(9)), gatekeeper compliance, APIs, Open Banking/financial data access

Luxury handbags: Prada Galleria (colors: emerald/deep blue), Louis Vuitton Capucines, Dior Lady Dior, Chanel Classic Flap; research into pre‑loved markets

Geopolitics and world affairs: North Korea, Russia, EU; documentaries/news (BBC, WSJ, Johnny Harris)

Tech founders and product: OpenAI/Anthropic/Cognition leadership interviews; a16z strategy; design icons (Jony Ive)

Destinations/planning: India (Delhi/Udaipur), Poland (Wrocław), Italy (Lake Iseo, Capri); UK leisure

With very little prompting, the agent picked up interactions that led to real purchases and it surfaced all of my upcoming trips without me pasting receipts or linking my calendar.

Q3 2024

AI, data, and personalization: recommendation systems, vector databases (Weaviate), hyper‑personalization, privacy impacts, context‑aware UX, A/B testing frameworks

E‑commerce stack: Shopify ecosystem (Rebuy, Braze, Axeptio), search/personalization vendors (Constructor, Typesense, Miros), receipts and fintech infra (Plaid, TrueLayer, Slip, ReceiptHero)

Founder/VC ecosystem: Demo Days (Entrepreneurs First), investor research (Forerunner, a16z, Northzone, Susa, 8VC, Floodgate, Headline, Red Swan, etc.) and operator talks (Zach Perret, Brian Chesky)

Luxury and contemporary: Prada Galleria, Louis Vuitton, Hermès Herbag, Tiffany, Van Cleef & Arpels

London/SF lifestyle: Third Space gym, Equinox; high‑end dining interest points (e.g., Chiltern Firehouse)

YOY Comparison (Q324 vs Q325)

The differences make sense.

  • Q3 2024: vector DBs, recsys, hyperpersonalisation, Shopify ecosystem, VC research, London/SF lifestyle

  • Q3 2025: heavier tilt to agents/MCPs, geopolitics, product research, European travel

The agent picked up the shift in my lifestyle and focus year over year, while also catching the steady interest in fashion.

Moment of truth: trial by avatar

For a laugh, I asked Nano Banana to generate an avatar from these summaries.

The agent has no photos of me. It just used my Google search and YouTube interactions to get here.

You can call the verdict.