Cobalt Capture for ChatGPT

This page is for anyone using ChatGPT who wants its research to end in something they can send to someone else. No terminal and no repo required. If you work in an editor, the rest of the feedback for AI coding agents hub covers Cursor, Claude Code, and the others.

The problem with an answer you cannot show anyone

Ask ChatGPT to look at a competitor's pricing page and it does the work well. It reads the tiers, runs the per-seat math, and tells you where the step up from Starter to Pro stops being defensible.

Then two things go wrong. The answer is text, because ChatGPT can read a page but cannot render one, so it will describe a cluttered pricing table without ever showing it to you. And the answer is stuck in a thread. When someone asks which plan you meant, or you want to re-run the same check next quarter, you are back to pasting links and explaining from memory.

What you wanted was a document: the verdict on top, the screenshots underneath, living at a URL.

The Cobalt workflow with ChatGPT

Custom connectors live behind Developer Mode, so it is two settings trips the first time and none after that.

1. Settings, then Apps & Connectors, then Advanced settings, and turn on Developer Mode.
2. Back in Apps & Connectors, choose Create.
3. Give it a name and paste the URL, including the path: https://cobaltcapture.com/mcp. Leaving the /mcp off is the single most common mistake.
4. Set authentication to None. There is no login and no API key.

Then ask in ordinary language:

• Run a pricing teardown of https://example.com/pricing
• Run a positioning audit of https://example.com
• Compare Linear, Jira, and Asana for a team of ten, with a Cobalt review

ChatGPT works through the study and returns a review URL. Open it and the summary sits at the top, each screenshot below with its finding written underneath, plus a save link that keeps it on your account.

Why this works for ChatGPT specifically

Cobalt takes the screenshots for it. ChatGPT has no browser it can capture from, so Cobalt renders each page itself and attaches the image, desktop and phone. The model sees what it captured and so do you. A teardown containing the actual pricing page is a different object from a paragraph about one.

The studies are written, not improvised. A pricing page teardown is a maintained playbook with a real method: a fixed question, consistent lenses, severity tied to outcomes rather than taste. Positioning audits and head to head comparisons work the same way. You supply a URL and confirm the plan once.

The result outlives the conversation. A review has its own link. Send it, keep it for the next re-run, or hand it to a developer who can act on it.

Alternatives and tradeoffs

Developer Mode is required to run studies, and it is the real gate. Without it ChatGPT only talks to connectors through a fixed pair of retrieval tools, search and fetch. Cobalt answers those, but they can only read, so nothing gets captured or written. Developer Mode is available on Plus and Pro; on Business, Enterprise, and Edu an admin can switch it off or allowlist connectors, and on Free there are no custom connectors at all.

Deep Research can read, not run studies. Deep Research and company knowledge reach a connector only through search and fetch. On Cobalt, search covers the studies we have published in the teardowns gallery, so a question like what Canva's Business plan really costs a team of five can cite the study, and fetch also reads any review link you paste in. It does not search your own reviews, which are private to their link. To run a new teardown, use normal chat with Developer Mode on.

Studies that need an account are limited here. A usability pass or a flow walkthrough has to sign up and click through a product, which needs an agent driving a real browser. From ChatGPT those cover the public, logged out screens and say so. Pricing teardowns, positioning audits, and comparisons read public pages, so they run start to finish.

Just asking without the connector is fine for a throwaway answer. You lose the screenshots and the shareable page, which is usually the part you wanted.