Point an AI at your site. Get the review back.
Pricing teardowns, usability passes, positioning audits: an AI runs the study and hands back a shareable review with real screenshots and the verdict on top. Readable by a person, actionable by your coding agent. Or capture and mark up a screen yourself; it lands in the same review.
Pick a study. Give it a URL. Get a review.
Each one is a researched method the AI follows step by step, delivered as a review you can share, export, or hand to a coding agent. The ones marked runs from chat need no browser at all: add the connector in Claude and ask in plain words.
-
Pricing teardown
runs from chatCan a buyer tell what they’d pay and which plan is theirs? The buyer math for three buyer sizes, tier names vs economics, the fine print, the phone view.
How it works -
Positioning audit
runs from chatWhat your site actually says to a cold visitor: who it’s for, what it claims, whether the page proves it. Desktop and phone, every key page.
How it works -
Usability pass
A first-time user, played by an AI with its own throwaway inbox: signs up, works the real jobs cold, screenshots every point of friction, ranks it by severity.
How it works -
Flow walkthrough
Name a flow, signup, checkout, inviting a teammate, and get back what broke and what merely confused, step by step, including any email it sent.
How it works
Visual feedback your coding agent can read
However a review gets made, it ends up in the same place: a shareable page with annotated screenshots and notes, exportable as markdown, PDF, Word, or a GitHub issue, and readable over MCP by Claude Code, Cursor, and Codex as real images, not descriptions.
Point it at a URL
Pick a study, hand over the address, and the AI does the work: it reads or drives the site, captures every screen it judges, and writes the verdict at the top. Re-run it later and the review leads with what got fixed.
See the studiesScreen, markup, voice
Capture a screen or paste a screenshot, drop pins and arrows, talk through what’s wrong. Publish and you have a link a developer, a client, or an agent can act on. No install, no signup to start.
Start a captureHow a capture works
Capture, narrate, hand off. The same loop whether the next reader is a developer, a client, a teammate, or an AI coding agent like Cursor or Claude Code.
-
1
Capture the screen
Click Capture screen, pick the window you want to review, drag a box around what matters. The screenshot is cropped and stamped with the source URL automatically. No annotation hell, no separate "save as" step.
-
2
Add a comment
Type your note, or hit Dictate to talk it through. Either way you get an editable paragraph next to each screenshot. Dictating tends to capture the context you'd skip when typing: what you noticed, a question, an idea, or what you'd change.
-
3
Publish and hand off
Hit Publish. You get a public review URL plus a markdown export. Paste the URL into your agent's prompt, or
curlthe markdown into your repo asfeedback.md.
One artifact, every receiver
Each published review is a public URL plus clean markdown, PDF, and Word exports. Polished docs for humans. Plain markdown for AI coding agents. Same review, every destination.
# Staging review, checkout flow
_Published Mar 14, 2026 via Cobalt Capture._
**Reviewing:** <https://staging.example.com/checkout>
**Source:** <https://staging.example.com/checkout>
The submit button overflows the container at viewport widths under
380px on iOS Safari. Should respect the existing max-width 100% rule
from .btn-primary.
---
**Source:** <https://staging.example.com/checkout>
Email validation fires on every keystroke. Should debounce or wait
for blur. The shake animation is also too aggressive for a
soft-validate.Built for the agent in your stack
Each page below is a dedicated playbook: the right prompt pattern, the exact paste destination, what the agent reads inline.
Same workflow, different jobs
The output works whether the reader is a developer, a designer, a client, or an AI agent.
-
Product review
PMs walking a build before launch, with feedback the dev (or their agent) can act on.
Read playbook -
QA bug reports
One artifact per bug. Cropped screenshot, source URL, dictated repro. No Slack-to-Linear ping-pong.
Read playbook -
Design review
Designers reviewing what dev shipped, on the built UI, with markdown the dev's agent can ingest.
Read playbook -
Client feedback
Send clients a public URL they can open without an account. Comment trail stays in your inbox, not a tool they wouldn't log into.
Read playbook -
Pull request review
Visual context for UI-touching PRs. The diff shows what changed in code, the review URL shows how it looks on every viewport.
Read playbook -
Competitive analysis
Walk a competitor's product once, narrate the patterns worth borrowing, hand the deck of screens to your team or your PRD.
Read playbook
Honest comparisons with the tools you have
Most feedback tools require a Chrome extension, a signup, or per-seat pricing before you can send one piece of feedback. Cobalt Capture runs in your browser, free.
Common questions
What is Cobalt Capture?
Cobalt Capture turns product feedback into a review an AI can act on, and an AI can produce. Point it at a site and it runs a study (a pricing teardown, a usability pass, a positioning audit) and hands back a shareable review with real screenshots and the verdict on top. Or capture a screen yourself, mark it up, and talk through it. Either way the result is a public URL plus markdown, PDF, Word, and MCP access, readable by a developer, a client, or a coding agent like Cursor or Claude Code. No install, free to try.
Which AI coding agents does Cobalt Capture work with?
Cursor, Claude Code, Windsurf, Lovable, Bolt, v0, Replit Agent, Cline, Aider, and Zed AI all read pasted markdown or .md files. Cobalt Capture produces both. Any agent that follows URLs or reads files in a repo will work. See all agent workflows for the per-tool playbook.
How is Cobalt Capture different from Loom?
Loom records a video. AI coding agents cannot watch videos. Cobalt Capture produces a markdown document with embedded screenshots, which every modern coding agent can ingest as input. Loom is still good for async demos to human audiences. See why Loom doesn't work with agents.
Do I need an account?
No. You can capture, narrate, and publish a review anonymously. Anonymous reviews are deleted after 30 days. Sign in with Google to keep reviews on your account and unlock editing.
Is Cobalt Capture free?
Yes, to try and to use: unlimited reviews, screenshots, voice dictation, and markdown export, no credit card. Free reviews stay live for 7 days and then lock; Pro ($5/month or $50/year) keeps every review on your account live. See the pricing page.
What does a Cobalt Capture markdown export look like?
A markdown file with the review title as the H1, an italic published-date line, an optional summary, and a section per finding with an embedded screenshot URL, the source page URL, and your dictated commentary as a paragraph. See markdown screenshots, explained.
Do I have to dictate, or can I type the commentary?
Both work. Every captured item has an editable text field you can type into directly. Dictation is the differentiator because voice tends to capture nuance you'd skip when typing, but it's optional. You can also dictate first and edit the transcript afterward.
Run your first study.
Give it a URL from Claude chat or Claude Code and get a shareable review back, verdict on top, real screenshots, readable by humans and coding agents alike. Or capture a screen yourself. No install, free to try.
Welcome to Cobalt Capture.
You're signed in, nothing here yet. Reviews you create will show up on this page with view counts and quick links to share. Start with your first one:
- Pop out. Cobalt Capture floats on top of whatever you're reviewing.
- Capture. One click per screenshot. Type or dictate commentary inline.
- Finish. Get a shareable URL plus markdown your agent can paste straight into a prompt.
Dashboard
Your recent reviews and how often they've been opened.
Recent reviews
All my reviewsSummary
Shown at the top of the public page. Markdown formatting is supported.
No captures yet. Click Pop out & capture above to start a floating capture session.
Already have a screenshot? Paste it with ⌘V / Ctrl+V, drop the file anywhere here, or use Add image.