You shipped a build with an AI coding agent. Nobody on the team does user research. You want to know where a first-time user gets stuck before a real one does, and you want the answer without scheduling a thing. Here are the questions founders actually ask before they run their first agent-driven usability pass, with direct answers under each.
Which AI agents can actually run a usability test on my site?
A usability pass needs an agent that can drive a real browser: sign up, click through tasks, and screenshot friction as it goes. Claude Code can, and so can Cursor, Codex, Windsurf, and the other agents that support a browser plus MCP connectors. The public-page studies (pricing, website, comparison) run in plain Claude chat with no browser, because Cobalt renders those pages itself. The usability pass is different. It operates your product, so it needs the browser.
If you want to see the full setup and what the agent does step by step, the page on having your agent run a usability test walks the whole thing. You connect Cobalt once, then ask in plain words.
How do I connect Cobalt to the agent I already use?
You add Cobalt as an MCP connector. In Claude chat: Settings, Connectors, Add custom connector, paste https://cobaltcapture.com/mcp. In Claude Code: claude mcp add --scope user --transport http cobalt https://cobaltcapture.com/mcp. Cursor, Codex, and Perplexity work the same way. Once it is connected, type /workflows to see every study and the exact phrase to trigger each one. There is no code to write and no prompt to engineer. You say something like "run a usability pass on my staging build" and the agent takes it from there. The MCP connection guide covers the one-time setup for each tool.
What do I actually type to start one?
Plain words. "Play a first-time user on example.com, try to sign up and complete the main task, and screenshot anything confusing." The agent signs up with a temporary test inbox, works through the real flow, and captures every point of friction as a still. You do not script the steps. You describe the goal and let it behave like a new user.
Will it sign up and poke at my product without asking me first?
No. The usability pass stops once, before it starts, for your go. That single confirmation exists because this study creates a real account on a product you own or are testing, which is different from the public-page studies that touch nothing. The agent tells you what it is about to do, waits, and only proceeds when you say yes. After that one moment it runs on its own. One go, then hands off. That handoff moment is the same idea covered in handing a half-built screen from human to agent, just pointed at a live flow instead of a static screen.
What is in the review when it finishes?
A shareable review, not a chat transcript you have to scroll back through. It leads with the verdict at the top, then lists findings ranked by what they cost you, each one tied to a real screenshot of the exact screen where the friction happened. A broken step in signup that loses users ranks above a label that is merely awkward. You open one link and read it in order.
The review lives at a short /r/<slug> link, with a clean read-only /s/<slug> page for anyone who just needs to skim. It also exports as PDF, Word, HTML, and plain markdown. If you want to hand the findings straight back to a coding agent, the markdown at /r/<slug>/markdown is the format an agent reads, screenshots included. A signed-in owner can send it to a GitHub repository as an issue. For a fuller breakdown of the sections, see what an agent-run study review contains.
How is this different from asking the agent in chat what it thinks of my site?
A chat review reads your pages and gives an opinion. A usability pass does the tasks. It finds the dead-end email confirmation, the button that does nothing on the third click, the form that rejects a valid input, things you only catch by acting like a user rather than reading like a critic. The difference between the two is laid out in an AI usability test versus a chat review. If your pages read fine but people still drop off, the usability pass is the one that finds out why.
What does it cost to try?
The review itself is free to produce. No install, no browser extension, no password. Cobalt asks for an email when a review is ready so it can send you the link; you sign in with Google or an emailed link. After 7 days a review goes read-only for you: editing, exports, Send to GitHub, and agent access stop, though the shared link keeps working for anyone you sent it to, comments and all, and nothing is deleted. Pro is $5 a month or $50 a year, keeps every review editable and exportable with no cutoff, reopens a read-only review instantly, and cancels any time. Full terms are on the pricing page.
What if my build blocks bots?
Because the agent drives a real browser the way a person would, it gets through flows that trip up crude scrapers. If you are testing a dev build and worried about bot blockers, that specific case is covered in testing a dev build with an agent. Connect Cobalt, point the agent at the flow, and read the ranked review it hands back.