Coding agents and MCP

UX Robot’s MCP server lets a coding agent, such as Claude Code, Codex or any client that supports the Model Context Protocol, work through the issues your journey reports found. The agent reads an issue, changes your code through your normal review, records the fix as shipped, and asks UX Robot to check it on the live website.

1. Find: list-ux-issues lists open issues across the organisations you can reach.
2. Read: get-ux-issue returns what to change, where, why, the proposed change and the evidence.
3. Fix: the agent edits your codebase and opens a pull request. UX Robot never touches your code or repository.
4. Ship: once the change is merged and live, mark-ux-issue-shipped records it with the pull request link.
5. Verify: request-ux-issue-verification creates a spot check. A browser-capable agent captures each area and records the verdict.

Coding agents use the same connection as desktop fieldwork. Open Settings → Desktop connections, add a connection for the computer, and add the instruction it shows to your agent’s MCP settings. The connection works across every organisation your UX Robot account can access and lasts 90 days; see Desktop connections.

The issue tools arrived in tool catalogue version 0.14.0. If your agent connected before that, reconnect the UX Robot server and start a new chat so it sees them. An agent acts with your account’s permissions: anyone who can view an organisation can list and read its issues, and marking issues shipped or requesting checks needs permission to edit the organisation.

Issue keys are written UXR-12; uxr-12 and 12 also work. Keys are numbered per organisation, so when the same key exists in two of your organisations, add organisation_id. Every tool returns a next_action telling the agent what to do next.

list-ux-issues

Lists issues, newest first.

• status (optional): open (needs a fix, the default), shipped, fixed, dismissed or all.
• organisation_id (optional): one organisation.
• page_url_contains (optional): only issues whose page address contains this text, such as /pricing.
• limit (optional): 1 to 50, default 20.

Returns organisations and issues, each with issue_key, organisation_id, status, title, page_url, where, last_check, ticket and found_on.

get-ux-issue

Returns one issue in full. Inputs: issue_key (required) and organisation_id (optional).

Returns the summary fields plus what_to_change, why_it_matters, impact, how_to_measure, proposed_change (steps from UX Robot’s after picture, such as the exact wording to use), done_when, evidence, journey (with the report link), issue_url and history. The history includes New wording is live once the page timeline finds the proposed wording on the live page, a good moment to call request-ux-issue-verification.

mark-ux-issue-shipped

Records that the fix is merged and live. Inputs: issue_key (required), reference (the pull request or commit link, or a short note) and organisation_id (optional). The issue becomes Shipped, waiting for a check and its history records who shipped it and from which agent. It refuses fixed and dismissed issues.

request-ux-issue-verification

Creates a spot check for up to 10 open or shipped issues from one organisation. Inputs: issue_keys (required, such as ["UXR-12", "UXR-14"]) and organisation_id (optional). Returns the spot check’s run_uuid. A request that overlaps a spot check already waiting returns that one instead of creating another.

A spot check is one task, claimed and submitted with the same tools as desktop fieldwork. Any connected agent that can control a browser can do it, for example Claude in Chrome; it does not have to be the agent that made the fix.

1. Call claim-analysis-autopilot-task with the run_uuid. The task lists each issue with its page, what was wrong, where, and the expected change.
2. Open each page on the live website, find the area and capture it. Record one observation per issue with evidence.issue_key set to the issue key, outcome pass when fixed and fail otherwise, and a description in observed_result.
3. Upload each screenshot: save a checkpoint with capture_pending_upload, then use prepare-analysis-autopilot-capture-upload and run its upload command.
4. Call submit-analysis-autopilot-task with result.issue_checks: one record per issue.

"result": {
  "summary": "Checked both issues on the live website.",
  "issue_checks": [
    { "issue_key": "UXR-12", "status": "fixed", "observation_ids": [881], "note": "Each card shows the phone number." },
    { "issue_key": "UXR-14", "status": "not_checked", "observation_ids": [], "note": "The page needs sign-in." }
  ]
}

status is fixed, partly_fixed, still_present or not_checked. Every verdict except not_checked must cite a captured observation from the task; not_checked needs the reason in note. UX Robot refuses a submission that leaves an issue out or cites no screenshot. A spot check writes no report and uses no analysis unit. The same result.issue_checks format is used when a journey run again checks earlier issues.

Paste one of these into Claude Code or Codex from your website’s repository.

Use UX Robot to list the open issues for our website. Fix UXR-12 in this codebase and open a pull request. Once it is merged and live, mark it shipped with the pull request link and request a verification.

List the UX Robot issues on pages containing /pricing, read each one, and propose the code changes as one pull request. Do not mark anything shipped until I confirm the deploy.

Claim the UX Robot spot check I just requested, check each issue on the live website, upload the screenshots and submit the verdicts.

• Agents can read issues, mark them shipped and request checks. They can never mark an issue fixed: only a check with a screenshot of the live page can.
• Agents cannot dismiss or reopen issues. Those decisions stay with people on the Issues page.
• Spot checks use only safe, reversible actions: no signing in, forms, purchases or contacting anyone. A page that needs sign-in is recorded as not checked.
• UX Robot never changes your code, repository or tickets. Fixes go through your own review, and linked tickets only receive comments.