Have your agent run a usability test
Point your coding agent at a URL and it signs up, works through the real tasks, and hands back a review of everything that got in the way.
Most feedback tools wait for a human to notice something. This runs the other way round: your coding agent opens your site as a brand-new user, tries to get something done, and reports back on everything that slowed it down.
It signs up for a real account, works through the jobs a real person would care about, screenshots every point of confusion, and hands you a shareable review with the verdict at the top.
No account to create. Put in your address, copy the block, paste it to your agent. It takes it from there — including walking you through setup if it needs to.
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.
What happens after you paste it
- It works out what your product is from your site, and proposes two or three people who'd realistically sign up, plus the jobs they'd be trying to finish.
- It stops and asks you. This is the only moment it needs you. If the personas are wrong, everything after them is confidently wrong — thirty seconds here is worth more than anything else you could do.
- It runs the sessions. Signs up with its own disposable inbox, reads its own verification email, and goes through each job cold — finding things the way a first-time user has to.
- It hands you a link. Headline verdict at the top, then every finding ordered by how badly it hurt, with the screenshot that proves it.
If you run it from inside your codebase, findings come back with the file and line to fix.
The more reliable setup
The line above works with no setup at all, but it depends on your agent fetching the playbook faithfully. Connecting the MCP server removes that step entirely — the playbook is delivered through the protocol, so there's nothing to fetch and nothing to summarize. If you're going to run this more than once, do it this way:
claude mcp add --scope user --transport http cobalt https://cobaltcapture.com/mcp
Restart Claude Code, and it becomes:
/mcp__cobalt__usability_pass https://example.com
Same playbook either way — the MCP version just saves you pasting a URL, and gives the agent direct tools for building the review as it goes.
What it won't do
It stops at anything irreversible. It never enters real card or identity details, never completes a payment, and when a flow hands off to Stripe or an OAuth screen it captures the handoff and stops there — the handoff is the finding, not what's behind it.
On a site you don't own, it asks you to confirm you've got a reason to be testing it before it creates anything.
Capture your first review.
About a minute from open tab to a shareable URL your agent can ingest.
Start capturing