Studies your AI can run for you
Ready-made site studies. Pick one, give it a URL, and an AI does the work — then hands back a shareable review with real screenshots, the verdict on top. You don't need to write code or know how to prompt it: each study below tells you exactly what to say.
From Claude chat, Cowork, or any assistant — no code
- Add Cobalt as a connector, once. In Claude: Settings → Connectors → Add custom connector → paste
https://cobaltcapture.com/mcp. No login, no key. (Other assistants: add it as an MCP server with that URL.) - Then just ask, in plain words. For example:
Run a pricing teardown of https://example.com/pricingRun a positioning audit of https://example.comCompare Linear, Jira, and Asana for a team of ten, with a Cobalt reviewWalk the checkout flow on https://example.com - Get the link. The assistant works through the study and hands back a review URL you can share, plus a save link to keep it on your account.
A chat assistant has no browser of its own, so Cobalt takes the screenshots for it — desktop and phone. Studies marked runs from chat below cover every page that way. Studies that must sign up or click through a product need a coding agent driving a browser; from chat they cover the public, logged-out screens and say so.
First-time-user usability pass
Point an agent at a URL. It signs up as a brand-new user, works through the real jobs someone would actually do, screenshots every point of friction, and hands back a review with the verdict on top — file-and-line root causes included if it's your own codebase.
What you need to do: It creates a test account using a throwaway inbox, so usually nothing from you. If the signup has a human-check, it tells you that up front rather than stalling.
In a chat assistant, say: Run a first-time-user usability pass on https://your-site.com
Using a coding agent instead (Claude Code, Codex, Cursor)?
Pick your agent, paste the block. If Cobalt isn't connected, the agent stops and gives you the exact command to run — rather than quietly doing a shallower version of the job.
Walk a flow you name
You say which journey matters — checkout, onboarding, inviting a teammate — and the agent walks it step by step, reporting what broke and what merely confused. Any email the flow sends is audited too, rendered and judged like any other screen.
What you need to do: Nothing, unless the flow is behind a login with a human-check. It stops before payments and anything irreversible, and says so before it starts.
In a chat assistant, say: Walk a specific flow on https://your-site.com
Using a coding agent instead (Claude Code, Codex, Cursor)?
Pick your agent, paste the block. If Cobalt isn't connected, the agent stops and gives you the exact command to run — rather than quietly doing a shallower version of the job.
Pricing-page teardown runs from chat
Reads your pricing page the way a buyer does: works out what a solo user, a team of five, and a company of twenty-five would actually pay, checks whether each tier's name matches its economics, whether the pushed tier is a nudge or a trap, what the trial and fine print really say, and whether any of it survives on a phone. Comes back with the price story your page transmits and the places a buyer goes wrong.
What you need to do: Nothing. It only reads public pages — no account, no checkout, no approval stop. Runs from Claude chat too: Cobalt renders the page itself.
In a chat assistant, say: Run a pricing-page teardown of https://your-site.com
Using a coding agent instead (Claude Code, Codex, Cursor)?
Pick your agent, paste the block. If Cobalt isn't connected, the agent stops and gives you the exact command to run — rather than quietly doing a shallower version of the job.
Comparison study runs from chat
Two or three products, one buying decision. It works out what decides the category first and proposes weighted criteria you can edit in a sentence, then captures each maker's page, researches every cell with a dated source, and comes back with a matrix that names the edge on every row, the critical differences, and an overall lean with what would flip it. Software tools or physical goods.
What you need to do: Nothing, beyond one reply when it shows you the criteria. It reads public pages and public reviews only; it never creates an account and it says so. Runs from Claude chat too: Cobalt renders the pages and the assistant does the searching.
In a chat assistant, say: Compare Product A and Product B for me, with a Cobalt review
Using a coding agent instead (Claude Code, Codex, Cursor)?
Pick your agent, paste the block. If Cobalt isn't connected, the agent stops and gives you the exact command to run — rather than quietly doing a shallower version of the job.
Positioning and messaging audit runs from chat
Reads your public pages the way a cold visitor does, on desktop and phone: who is this for, what does it claim, does the page prove it. Comes back with the positioning your site actually transmits, next to what you meant it to.
What you need to do: Nothing at all. It only reads public pages — no account, no signup, no approval stop, so it runs start to finish on its own.
In a chat assistant, say: Run a positioning and messaging audit of https://your-site.com
Using a coding agent instead (Claude Code, Codex, Cursor)?
Pick your agent, paste the block. If Cobalt isn't connected, the agent stops and gives you the exact command to run — rather than quietly doing a shallower version of the job.
Something not listed?
Ask for it anyway. The assistant won't force your question into the nearest study — it runs the same method every study here uses (state the question, capture every screen it judges, rank findings by what they'd cost the person the study is about, keep-list what works) and tells you it was a custom run. If you find yourself asking for the same custom study twice, tell us — that's how the next ready-made one gets built.
Cursor and VS Code install in one click. Claude Code and Codex copy a command to run in your terminal, then restart the app.