AI coding assistants (MCP)
Let Claude Code, Cursor or another coding assistant read a Midge report's console, network requests and steps over MCP.
Midge has an MCP server at https://midge.dev/mcp. Add it to Claude Code, Cursor or any assistant that supports remote MCP servers, then paste a report link and ask it to find the cause: the assistant reads the console errors, the failed requests with their bodies and the user’s steps itself. It gets the same redacted data you see on the report page, and it can only read.
Connect an assistant
Claude Code: run this in your terminal.
claude mcp add --transport http midge https://midge.dev/mcp
The first time Claude Code uses Midge, it opens midge.dev in your browser. Sign in if you need to, check the assistant’s name and click Allow. You’re sent back to Claude Code, signed in. (You can also start it from Claude Code with /mcp.)
Cursor: add the server to ~/.cursor/mcp.json, or to .cursor/mcp.json in a project. Cursor signs you in the same way.
{
"mcpServers": {
"midge": { "url": "https://midge.dev/mcp" }
}
}
Other assistants: add a Streamable HTTP server at https://midge.dev/mcp. Clients that support MCP sign-in find the rest on their own.
With a token instead
For an assistant that can’t sign in from the browser, or for a script:
- On midge.dev, open Settings › Account.
- Under AI assistants, pick Claude Code, Cursor or Other, and click Create token.
- Copy what Midge shows; the token appears only once. It sends the token as the header
Authorization: Bearer <your token>.
On a report
Reports with logs have a Fix with AI button next to Copy link. It copies a prompt with the report’s link for your agent, and shows the command to connect one.
Use it
Paste a report link and say what you want, for example: “Here’s the bug report: https://midge.dev/r/… Find why checkout fails.” The assistant chooses among these tools:
| Tool | Reads |
|---|---|
get_recording | The overview: title, description, page, device, duration, the pages visited, and the first console errors and failed requests. |
get_console | Console messages in order, with stack traces. Filter by level (error, warn, all) or text. |
get_network | Requests with method, status, type and timing. Ask for failed requests only, one endpoint, and headers and bodies (cut at 2,000 characters). |
get_user_actions | Clicks, typing (never the text), scrolls and page changes, in order. |
list_recordings | Your recent reports, newest first, with a search. |
Long lists come in pages; the assistant asks for the next one when it needs it. Times are counted from the start of the recording, as on the report page.
Access and privacy
- The assistant can open what you can open: your workspaces’ reports (not someone else’s “Only me” reports), and any report shared as Anyone with the link when it has the link.
- It reads only. It can’t change, delete, comment or share.
- Everything it reads was redacted before upload, in the browser or on your Mac. See Redaction.
- Each assistant is listed under Connected devices as “Claude Code (MCP)”, “Cursor (MCP)” and so on, with when it was last used. Click Sign out to stop it at once.
- Up to 120 tool calls a minute per person.