Skip to content
Sign inRequest access

MCP server (AI agents)

uMarker ships an MCP (Model Context Protocol) server, so an AI coding agent can pull the context it needs to reproduce and fix a bug: the reporter’s words plus browser metadata, the report’s labels and any linked issue — and, on a report filed in Technical capture mode, console logs, network requests, errors and breadcrumbs. It can then mark the report resolved and leave a comment when the fix ships.

The server is a thin wrapper around the REST API: it runs on your own machine, speaks JSON-RPC over stdio, has no dependencies, and carries your API key on every call. uMarker never talks to your agent. Your agent talks to uMarker.

  1. Mint an API key: Dashboard → API Keys. Account-wide, or scoped to the one project the agent should see. Copy the md_live_… secret; it is shown once.

  2. Get the server file. It is one file, umarker-mcp.mjs — about 6 KB of plain Node with no dependencies and nothing to install (Node 18 or newer). It is not on npm and there is no download button on this site yet: ask for it at support@umarker.com, or from Dashboard → Support. Save it anywhere; the example below assumes ~/umarker-mcp.mjs.

  3. Point your agent at it. For Claude Code:

    نافذة طرفيّة
    claude mcp add umarker \
    -e UMARKER_URL=https://api.umarker.com \
    -e UMARKER_API_KEY=md_live_your_key_here \
    -- node ~/umarker-mcp.mjs

    For any other MCP client, configure a stdio server with the same command and environment.

Variable Value
UMARKER_URL https://api.umarker.com: the API base. The server appends /api/v1 itself.
UMARKER_API_KEY An md_live_… key from Dashboard → API Keys.
Tool What it does
list_projects The projects your key can see: how an agent resolves a project id
list_reports Reports in a project, newest first (per_page default 25, capped at 100 — and raised to 5 if you ask for fewer)
get_report Full context for one report: console, network failures, metadata, labels, linked issue, plus a paste-ready markdown block
update_report_status Move a report to new, in_progress, waiting, resolved or archived
add_comment Leave a note on the report (fix summary, PR link), max 5000 characters

Both write tools raise the same events the product raises, so nobody has to watch the API to learn what the agent did. A status change fires status_change — or resolved, which is its own event whenever the new status is resolved — and either one reaches the team’s bell, their email, and any distribution route subscribed to it.

A comment fires comment, which reaches the bell and the email but never a channel: comment is not one of the seven events a route can subscribe to. So if the agent’s “deployed, please verify” has to show up in Slack, have it resolve the report as well as comment. Moving a report to resolved is also what opens the team’s close-the-loop options.

You: fix the top bug in uMarker

Agent: calls list_reports, picks #42 “Checkout button does nothing”, calls get_report, sees TypeError: cart is null in the console and a failed POST /api/cart (500), finds the handler, fixes the null path, runs tests, then calls update_report_status(42, "resolved") and add_comment(42, "Fixed in PR #128: cart was read before init on cached pages.")

No copy-pasting stack traces, no “can you send a screenshot of the console?”.

Every report page has a Copy report as text button that puts the same context block on your clipboard. Paste it into any chat or agent, MCP configured or not. It needs the feedback.export permission, because lifting a report out of the product is its own decision.

  • The MCP server runs on your machine and talks only to uMarker, with your key. Nothing is proxied through us and no third party sits in between.
  • Keys are shown once, stored hashed, revocable, and scoped (account or single project). See API keys.
  • Whatever was redacted was redacted on the reporter’s device, not here. Query-string values whose names look like secrets are blanked on every project, so a token in a failed-request URL never reached uMarker and cannot reach your agent. Field values the visitor typed and the regions you mark with a CSS selector are masked the same way — in the reporter’s own browser, before the screenshot is uploaded — but only on a project with Auto-redact sensitive fields switched on, and a new project ships with it off. Until it is on, the screenshot and, on a Technical report, the console lines beside it carry whatever was on the page, and get_report hands them to your agent. See PII redaction.
  • A report older than your plan’s history window is invisible here too: the API and the MCP server honour the same window the dashboard does.