You can run a website study, get back a review full of screenshots and ranked findings, and have Cursor read every item without you copying a single line. The link between the two is the MCP connector. Here is the full path, step by step, from an empty chat to Cursor working off the review.
Step 1: Add Cobalt to the assistant you run the study in
The study runs inside an AI assistant, not in Cursor. Start in Claude chat or Perplexity, whichever you already use. In Claude chat, open Settings, then Connectors, then Add custom connector, and paste https://cobaltcapture.com/mcp. If you prefer the command line, Claude Code takes claude mcp add --scope user --transport http cobalt https://cobaltcapture.com/mcp. The result: your assistant can now run Cobalt's studies and publish reviews.
The same connector string works across any assistant that supports MCP. If you want the broader picture of what MCP is and how the connection behaves, the guide on connecting your coding agent over MCP covers it without the Cursor specifics.
Step 2: Ask for a study in plain words
Type what you want. For a landing page, something like "Run a website teardown of yoursite.com" is enough. For a public pricing page, "Run a pricing teardown of yoursite.com/pricing". No code, no prompt engineering. The public-page studies (pricing, website, comparison) run start to finish with nothing for you to do, because Cobalt renders the pages itself instead of driving a browser. Type /workflows first if you want the full list and the exact phrasing for each one.
The outcome of this step: a published review at a short link of the form /r/<slug>. It leads with the verdict, then real screenshots of each page judged, then findings ranked by what they cost you. That link is the handoff point.
Step 3: Read the review yourself before anything touches Cursor
Open the /r/<slug> link in a browser and skim it. You are deciding which findings are worth a code change and which are noise you disagree with. Cursor will act on what you point it at, so the triage is yours to do, not the agent's. People reviewing the same build on different screens can also open the clean /s/<slug> read page, which strips out the editing controls.
If you want to avoid the common ways this handoff goes sideways, the notes on feedback that lands badly in Cursor are worth two minutes here.
Step 4: Point Cursor at the review over MCP
Now add the same connector to Cursor. Cursor supports MCP connectors, so it reads from Cobalt the way Claude does. Once connected, you ask Cursor to read the review by its slug or its URL, and it pulls the markdown version at /r/<slug>/markdown automatically. Screenshots come through too, not just the text, so the agent sees the pinned spots and the cropped stills, not a vague description of them.
This is the part that removes the copy-paste. You are not pasting screenshots into a chat box or retyping findings as a to-do list. Cursor reads the artifact directly. The full setup, with the Cursor-specific details, is in the guide on getting feedback into Cursor, and there is a shorter companion piece on pulling a review into Cursor without copy-paste if you only need the connection part.
Step 5: Tell Cursor what to do with the findings
With the review loaded, give Cursor the scope. "Fix the top three findings" or "Turn each item into a task and start with the pricing table one" both work, because the agent now has the ranked list in front of it. Because findings are ordered by cost, "start at the top" is a real instruction, not a guess.
The agent reads each pinned screenshot and the comment attached to it, so a note like "this button is below the fold on mobile" lands against the actual pixel it was pinned to. That is the difference between an agent editing the right element and one editing something nearby. If you want to understand why the screenshot itself matters more than a link to the live page, see why your agent should read the screenshot.
What this costs and when it stops working
Running studies and reading reviews is free to try, with no password and no credit card. Cobalt asks for an email when a review is ready so it can send the link. The one limit worth knowing before you build this into a routine: after 7 days a review goes read-only for its owner, which stops editing, exports, and agent access to it. The shared link keeps working for the people you sent it to, and nothing is deleted. Pro at $5 a month, or $50 a year, keeps every review editable and agent-readable with no cutoff and reopens read-only ones instantly. If you plan to re-run a study after a redesign and pull the new review back into Cursor, that cutoff is the thing to weigh. The full breakdown is on the pricing page.
That is the whole loop: run the study in one assistant, read the ranked review yourself, connect Cursor over the same MCP server, and let it work off the markdown and the screenshots directly. The next build, you re-run the study, and the review leads with what got fixed.