Connect Claude and ChatGPT to Your Research: MarklyKit's New MCP Connectors

For two years, the workflow for "asking an AI about my research" has been the same depressing dance: open ChatGPT or Claude, paste a wall of text, hope the model has room for it, hope you didn't forget the one paragraph that mattered. The AI is brilliant; the plumbing is medieval.

MarklyKit Pro now ships a built-in MCP (Model Context Protocol) server. With one click in Claude.ai — or a custom connector in ChatGPT — your highlights, sticky notes, folders, and tags become first-class tools the AI can search, filter, and pull from on demand. No exports. No pasting. No "let me find that quote again."

This post explains what MCP is, why it matters for research, and exactly how to set it up.

What Is MCP, in One Paragraph

Model Context Protocol is an open standard (introduced by Anthropic, now supported by OpenAI and others) that lets AI assistants talk to external tools the same way a developer's IDE talks to a language server. Instead of you pasting context into a chat, the AI calls a tool — search_research, get_folder_research, get_highlights_for_url — and pulls only what it needs, when it needs it. Think of it as giving the model a librarian instead of dumping the library on its desk.

For a refresher on why structured research beats raw notes pasted into a chatbot, see our earlier piece on taking better research notes while browsing the web.

What Claude and ChatGPT Can Now Do With Your MarklyKit Research

Once you've connected MarklyKit, the AI gets a small set of tools. The interesting part isn't the tools themselves — it's the questions they unlock:

• "Summarize everything I've highlighted on retrieval-augmented generation this month." The model calls search_research with your query, gets back the relevant highlights with their source URLs, and writes a coherent summary with citations.
• "What did I save in my 'Thesis — Lit Review' folder?" The model calls get_folder_research, pulls every annotation in that folder, and structures it however you ask — outline, table, bullet list.
• "I'm reading this Wikipedia article. What did I already note on it?" get_highlights_for_url returns your past annotations for that exact page so you don't repeat work.
• "Find the contradictions across my sources on X." The model searches, reads, and cross-references — something that's genuinely tedious for a human.

These aren't hypotheticals. They're the exact prompts our beta users have been running for the past month.

Why This Beats Copy-Paste

There's a real qualitative shift here, not just a convenience one.

The model only loads what's relevant. Pasting your entire research notebook into a chat blows out the context window and dilutes the signal. MCP lets the model retrieve precisely — three highlights, not three hundred.

Citations come for free. Every result includes the source URL, the highlight text, your note, and the surrounding context. Ask for a summary and you'll get inline links back to your original sources. (This is also why we built semantic anchoring — highlights need to be reliably attached to the underlying text for citations to survive.)

Your folders become AI lenses. If you've organized your reading into folders ("Competitive Analysis," "Q2 OKRs," "Dissertation Sources"), each folder is now a filterable corpus the AI can reason over in isolation.

Your research stays yours. The MCP server runs against your private MarklyKit cloud. Only highlights, folders, and sticky notes are mirrored to the cloud — drawings and blurs stay device-local. Only your authorized AI clients can read it, and you can revoke access at any time from Settings.

How to Connect (3 Minutes)

Claude.ai (Pro / Max / Team)

1. Open Claude.ai → Settings → Connectors → Browse connectors → Add custom connector.
2. Paste your MarklyKit MCP URL (you'll find it on your Settings page under "Connect Claude.ai or ChatGPT").
3. Click Connect. Claude opens MarklyKit in a new tab, you authorize once via OAuth, and you're done.
4. Open the connector inside Claude and set each tool to "Always allow" so it can call them without re-prompting.

ChatGPT (Plus / Pro)

1. In ChatGPT, open Settings → Apps & Connectors → Advanced settings and enable Developer mode.
2. Back in Apps & Connectors, click Create to add a custom connector.
3. Name it MarklyKit, paste your MCP server URL, set Authentication to OAuth, and click Create.
4. In a new chat, open the + menu → Developer mode, enable MarklyKit, and call its tools (e.g. @MarklyKit summarize my latest highlights).

Developer mode is the only way to add custom MCP connectors in ChatGPT today — it requires a Plus or Pro plan.

Five Real Workflows We've Already Seen

1. Literature review draft. A PhD candidate asked Claude to "outline the open questions across everything in my Thesis folder." Claude pulled 80+ annotations, grouped them by theme, and flagged three contradictions she'd missed.
2. Competitor digest. A founder runs a weekly prompt: "What's new in my 'Competitive Intel' folder since last Tuesday?" Five minutes of reading replaces an hour of tab-juggling. (We covered the broader research-workflow story in 5 Ways MarklyKit Transforms Your Online Research Workflow.)
3. Article-level recall. Reading a long technical post? Ask "what did I highlight on this URL last time?" before you start — saves you from re-highlighting the same passage twice.
4. Cross-source synthesis. "Compare the arguments in the three Stratechery posts I tagged with platforms." The AI fetches only those, then writes the comparison.
5. Studying for exams. Students with hundreds of highlights across textbooks and articles ask the AI to quiz them on a folder. The AI sees the source material; you see the questions.

If you're new to web annotation in general, our complete guide on how to highlight and annotate any webpage is the right place to start — the AI workflows above only work if you've actually captured something worth querying.

A Note on Privacy

MarklyKit is offline-first. Annotations write to your device first, then sync in the background on Pro. Only highlights, sticky notes, and folders sync to the cloud — exactly the data the AI tools need to be useful. Freehand drawings and blur regions stay on-device because they're tied to a page's exact layout and don't travel well between devices anyway.

We don't train on your data, we don't sell it, and the OAuth flow means you authorize Claude or ChatGPT — not us — to read your research. Revoking access is one click in Settings.

Why We Built This

We've written before about web annotations being a secret weapon for serious researchers. What's changed in 2026 is that "serious research" increasingly includes an AI in the loop — not as a writer, but as a faster reader. The bottleneck isn't generating text; it's pulling the right context together.

MCP is the first protocol that makes this clean. Your tools talk to the AI on equal terms. Your data stays in one place. The AI is just another reader who happens to be very fast at synthesis.

If you've been treating ChatGPT or Claude as a glorified text box, this is the upgrade. And if you're still curating your toolkit, our roundup of the best Chrome extensions for students and researchers in 2026 is a good companion read — MarklyKit was always built to slot into a serious research stack, not replace one.

Get Started

The MCP connectors are part of MarklyKit Pro ($7.99/mo). Highlighting, sticky notes, and folders work on the free plan; cloud sync, AI connectors, semantic search, and PDF exports unlock with Pro.

Add MarklyKit to Chrome — Free → — start highlighting, then upgrade to plug Claude or ChatGPT into your research when you're ready.