Midge
Lets Claude Code, Cursor and other agents read Midge bug reports: the screen recording or screenshot, console logs, failed network requests and user actions, with secrets already redacted. Remote server at https://midge.dev/mcp with browser sign-in (OAuth).
Hosted MCP Server
npx add-mcp 'https://midge.dev/mcp'Installs into Claude Code, Codex, Cursor and more
Documentation
A bug report MCP server so your coding agent can read the report
Midge has a bug report MCP server. It lets coding agents such as Claude Code and Cursor read a Midge report directly: the console errors, the network requests and responses, the user's clicks and page changes, and the browser details. Instead of you copying a stack trace into a chat, you paste the report link and the agent fetches the parts it needs.
Set it up
In Claude Code, one command:
claude mcp add --transport http midge https://midge.dev/mcp
The first time Claude Code uses it, your browser opens midge.dev: sign in, click Allow, and you're back in the terminal. Cursor and other clients that support remote MCP servers work the same way with the address https://midge.dev/mcp.
An agent that can't sign in from the browser uses a token instead: on midge.dev, open Settings › Account › AI assistants, pick the agent and click Create token, then send it as Authorization: Bearer <token>.
Either way, the agent shows under Connected devices as "Claude Code (MCP)" or similar; sign it out there and it can't read anything anymore. The docs page has the details for each client.
The tools
- get_recording: the place to start with a link. The title, description, page, device and duration, the pages visited, and the first console errors and failed requests.
- get_console: messages in order, by level (errors only, warnings and errors, or everything), with stack traces and a text search.
- get_network: requests with method, status, type and timing. Ask for failed requests only or one endpoint, and for headers and bodies when the agent needs them, so it reads what matters instead of a 4 MB log.
- get_user_actions: clicks, typing (never the text), scrolls and page changes in order, ready to become steps to reproduce.
- list_recordings: your recent reports, with a search, for when you didn't paste a link.
A typical session: you paste a report link into Claude Code and ask it to find the cause. The agent reads the overview, asks for the failed requests with their bodies, finds a 500 from POST /api/orders with a validation error, reads the matching client-side error in the console, and opens the handler in your repository.
What MCP is
The Model Context Protocol is an open standard that lets AI applications connect to outside tools and data. A server exposes a set of tools, such as "get the console log of report X", and a client like Claude Code or Cursor calls them while working on a task. You connect a server once, and the agent decides when to use it.
For bug fixing, this matters because the useful evidence lives in the bug report, not in the code. An agent that can read the failing request and the exact error can go straight to the right file.
Access and privacy
The server uses the same access checks as the rest of Midge. An agent with your token can open what you can open, and nothing more: a report set to "Only me" stays yours, and a workspace report stays with its members. Anyone's report shared as "Anyone with the link" opens for any signed-in agent that has the link, as it does in a browser. The server only reads; it can't change, delete or share anything.
Reports are redacted before they ever reach Midge. Passwords, card numbers, tokens and cookies are removed on your device before anything leaves it. The agent receives the same redacted data a teammate would see. The rules are listed on the redaction page.
Without MCP
If your agent doesn't speak MCP, you can still move a report into a chat by hand:
- Copy as cURL: open the failing request in the network logs panel and copy it. The
Authorizationheader and cookies arrive as[REDACTED], which is what you want in a chat transcript. - Download the redacted HAR: attach the report's network log as a HAR file. Agents handle it well; to look at it yourself first, open it in the free HAR viewer.
- Paste the console: copy the error and its stack trace, and the warning just before it if there is one.
Why MCP
We could have built a plugin for one editor. MCP is supported by many agents and editors, so one server serves Claude Code, Cursor and whatever your team uses next. It also keeps the agent in control of what it reads: it asks for exactly the requests it needs.
The Jira, Linear, GitHub Issues and Slack integrations are next.
Midge is free: unlimited recordings, videos up to 15 minutes, and no watermark.
Frequently asked questions
Is the Midge MCP server available now?
Yes, and it's free. Add https://midge.dev/mcp to your agent (in Claude Code: claude mcp add --transport http midge https://midge.dev/mcp) and sign in to Midge in the browser when it asks. Agents that can't sign in use a token from Settings › Account › AI assistants.
Which agents does it work with?
Any client that supports remote MCP servers over Streamable HTTP with a header, including Claude Code and Cursor.
What can the agent read?
A report's overview (title, description, page, device, pages visited), its console messages, its network requests with filters and bodies, and the user's steps. It can only read, and only reports that the person who made the token can open.
Will the agent see my secrets?
No. Reports are redacted before upload, in the browser or on your Mac, so Midge never stores the raw values. The MCP server can only return what Midge has, which is the redacted data.