UserProof

Point to the file and line to change: what it does and how to use it

Updated . By Ryan Doyle, founder of UserProof, a session replay tool for coding agents. For founders who build with a coding agent.

UserProof names the broken element, and your coding agent finds it in your repo. In Claude Code, the userproof-locate script returns a file and line for a scan defect. In our test on 40 elements from our own site, it found the right file 40 times and the right line 34 times.

This page is for founders who build with a coding agent (Claude Code, Codex or Cursor) and want a replay finding to end at a line of code. A stack trace can do that for a thrown error, but a dead click throws nothing, so there is no trace to read. UserProof's server has no copy of your code. It sends the page, the element's text and its CSS selector, and the search runs on your machine.

Jump to: which route you get · what it does · use it · our test · limits · price · sources · FAQ

Which route finds your code?

It depends on what produced the finding and which agent you run. Each row is what happens today, read from the plugin and server source on November 10, 2026.

Your situationHow the code gets foundWhat you get back
A scan defect in Claude CodeThe agent saves the defect as JSON and runs userproof-locateA best file and line, a score, the reason, and up to 5 candidates
A recorded-visit finding in Claude CodeThe review skill tells the agent to search this repo by page, element text and selectorThe file and element, then the change as a diff
Codex, Cursor or another MCP agentThe reply names the element; the agent uses its own search. The locator ships with the Claude Code pluginWhatever the agent's search finds
The locator returns nothingThe setup skill tells the agent to search for the element text and the page's route fileA file, or a plain statement that it was not found
A page-level finding (visitors bounce from one source, or never reach the button)No element to search for, so the agent opens that page's file and judges itA hypothesis about the page, not a matched line

What does it do?

It turns the description of an element into ranked lines in your repo, and your agent then opens the best one to confirm it. The finding comes from recorded dead and rage clicks or from a scan of the live site, and it says which element failed.

The locator takes that element and scores every line of your source files. It lists files with git ls-files, makes no network call, and has no dependencies. Each line gets its strongest match, plus points when the file belongs to the page:

Match on the linePoints
The line that threw (load errors only)90
Exact element text60
Element id, link href, or the whole class list55
aria-label50
Element text in any case45
File matches the page's route+40 (+15 for a partial path match)
The right tag on or just above the line+10

Text that came from a title or aria-label counts for half, because it is often data rather than a string in your source. A file under components, app, src, pages, routes, views or layouts gets 5 more points, and a markdown file loses 20.

How do I use it?

Ask your agent for the top fix, let it find the code, and read the file before you accept a change. Setup is on the Claude Code page and the Codex and Cursor page, so these steps start after it.

  1. Get one finding. Prompt: Call get_top_fix for this site and find the code it names. Do not change anything yet. For a site with no visitors, say scan_site instead. Done looks like a reply starting TOP FIX on or SCAN RESULT for, with a finding or defect id.
  2. Find the line. For a scan defect in Claude Code, the agent runs the command the reply names:
    node "${CLAUDE_PLUGIN_ROOT}/bin/userproof-locate.mjs" --repo . --defect /tmp/userproof-defect.json
    Done looks like a file and line in best, or an empty result the agent says out loud.
  3. Confirm it in the file. The skill tells the agent to open the file and check the element is really broken: no handler on a clickable element, a slow await, a link to a route that does not exist. If the code shows the finding is wrong, it says so and stops. Done looks like one sentence on why this line is the cause.
  4. Review the diff, then ship it. You get the fix in one line, the sessions behind it and the exact change. Done looks like your commit, live in production.
  5. Check it worked. The agent calls mark_fix_applied, and check_fix compares before and after once about 30 human visits have hit the page. A scan defect is re-tested by repeating the scan test. See check whether a shipped fix worked.

How accurate is it? Our test

On 40 elements from our own site it picked the right file every time and the right line 34 times (85%). We ran it on November 10, 2026 against this site's own static HTML: a random sample of buttons and links from 9 pages, passing each element's text, tag, class and href, and comparing the answer with the line where the element really is.

ResultElementsNote
Right file and right line34 of 40The answer an agent can open directly
Right file, wrong line6 of 40All six were text repeated in that file: 4 "UserProof" links, 1 "Copy" button, 1 "pricing page" link
Wrong file0 of 40Each element's text was on its own page

Here is one real answer, for the "Add a step" button on our funnel calculator, cut to the fields that matter:

"best": {
  "file": "funnel-calculator.html",
  "line": 87,
  "score": 110,
  "why": "exact text, route match +40"
}

A separate run shows the failure to watch for. We asked for "Start free" on /pricing, text that page does not contain. The locator returned line 92 of a benchmark page, inside its JSON-LD, with a score of 60 and no route match. A score with no "route match" in why is a hint at best, which is why the agent opens the file before it edits anything.

This was one repo of plain HTML, the easy case. We have not measured it on a React or Vue codebase, where text often lives in data files.

What can it not do?

It finds text in source files, so anything that is not written in the source can be missed or matched wrongly.

What does it cost?

It costs nothing extra: the script ships with the plugin and runs on your machine, and analyses are unlimited on every plan.

PlanPriceRecorded visits a monthSitesFinding the code
Free$05,0003Included
Pro$29 a month50,000UnlimitedIncluded
Scale$99 a month250,000UnlimitedIncluded

Prices are from the pricing page, read November 10, 2026. Comparing tools rather than setting this one up? Session replay tools for coding agents, ranked covers what each one does.

Sources, read November 10, 2026

FAQ

Does UserProof upload my code to find the line?

No. The server sends the page, element text and selector to your agent, and the search runs on your machine. The locator script lists files with git, reads them locally and makes no network call.

Does the locator work in Codex and Cursor?

The script ships with the Claude Code plugin, so Codex and Cursor do not run it through the plugin. They receive the same element details from the MCP tools and find the code with their own search.

How often is the file and line right?

On 40 elements from our own static HTML site on November 10, 2026, the file was right 40 times and the line 34 times. All six line misses were text repeated in the same file. We have not measured it on a React or Vue codebase.

What if the locator points to the wrong place?

The agent is told to open the file and confirm that it renders the element before changing anything. If the locator returns nothing, it searches for the element text and the page's route file, and if the code shows the finding is wrong, it says so and stops.