Website feedback your coding agent can act on
An AI tool can now produce a working site in an afternoon. Reviewing it
still takes a week, because feedback arrives as a screenshot with a red
circle, a Slack message that says "the pricing bit feels off", and a
comment about a version you replaced two builds ago. None of it tells
the person — or the agent — doing the fixing where.
Comma's Websites feature fixes the pointing. Upload the build, send
the preview, and every comment lands on the exact text it is about, on
the exact page, against the exact deployment.
How it works
- Build the site the way you already do —
npm run build, hugo,
astro build, whatever produces a folder of HTML, CSS and JS.
- Upload it on the Websites tab. A zip with
index.html at the root
works, and so does a project zip that contains a dist/, build/ or
out/ folder — Comma finds the output.
- Comma hosts it on its own isolated deployment. Your site's code
never runs on Comma's origin, so it can't reach reviewers' sessions.
- Invite reviewers. They click through the real site, select a
heading, a button label or a paragraph, and leave a comment. The
highlight stays on the text for the next reader.
- Hand the list to your agent — or work through it yourself.
What a comment carries
The reviewer sees a clean thread: their note, the quoted text, replies,
resolved or open. Underneath, each comment also keeps what someone
fixing it needs:
| The reviewer sees |
Also stored with the comment |
| Their comment and replies |
The page path it was left on |
| The highlighted text |
A selector for the element |
| Open / resolved |
The deployment (the exact build) it was on |
That is the difference between "the button is wrong" and "on
/pricing, the Buy now button inside the second card, on build 7".
Close the loop with Claude Code or Cursor
The comment list is available over Comma's MCP server and the
REST API:
list_websites and read_website find the site and its live build.
list_website_comments returns every comment with its page, element
and deployment context.
set_website_comment_status marks a comment resolved.
So the loop becomes: reviewers comment, you ask your agent to "work
through the open comments on the pricing site", it edits the code and
resolves each item as it goes, you upload the new build. Reviewers see
their threads resolved instead of being asked whether it's done. Setup
takes a minute in Claude Code or
Cursor.
Where it fits against feedback widgets
Tools like Marker.io, BugHerd and Pastel are good at what they are built
for: a widget on a site you already host, feeding screenshots into Jira
or Trello for a person to triage. If that is your process, keep it.
Comma is built for a different shape of work:
- You don't have a host yet. The build is on your laptop or came out
of v0, Lovable or Bolt. Comma hosts the preview for you — no Vercel
project, no staging environment, no DNS.
- An agent does the fixing. Comments come with the context a coding
agent needs and can be resolved by the agent, not re-typed into a
ticket first.
- The site sits next to your reports. The same workspace holds your
HTML reports, test results and AI-generated pages, with the same
comment model.
Honest limits
- Reviewers sign in. A Website is private by default; invite
commenters or make it public with comments on. To share a single screen
more widely, publish it as a Comma report instead —
anyone with the link can read it, and commenting takes a free sign-in.
- Uploads carry text files. Images and fonts should use absolute URLs
for now.
- Unbuilt source needs Vercel. Built output always works; to deploy
raw Next.js or Vite source, connect a Vercel account and it builds
there.
- Previews expire. Deployments are review builds, not production
hosting. Upload again when you want a fresh one.
- Linked sites aren't annotated. You can attach an existing URL and
keep page-level comments beside it, but in-page highlights need a build
Comma serves.
Try it
Upload a build to Comma →
Related