A user files a bug. The report contains 300 console entries, 150 network requests, a session replay, environment details, and a screenshot. Someone still has to read all of that, correlate the failed request with the console error, and map it to a line of code.
That reading job is exactly what AI coding assistants are good at — if they can reach the data. BugDump ships an MCP server so that Claude Code, Cursor, Claude Desktop, or any MCP-capable client can page through your bug reports directly and investigate them in natural language.
What MCP is, briefly
MCP (Model Context Protocol) is an open standard for connecting AI assistants to external data sources and tools. Instead of copy-pasting logs into a chat window, you point your assistant at an MCP server, and it can query that data itself — deciding what to fetch, in what order, as it works through a problem.
For bug reports, this matters because the useful signal is scattered. The console error is one clue, the 500 response two seconds earlier is another, and the replay showing the user double-clicking a submit button is a third. An assistant with MCP access can pull each piece as its hypothesis develops, the same way you would flip between dashboard tabs.
What the BugDump MCP server exposes
Once connected to a project, your assistant can read the full context of its bugs:
the reports themselves — descriptions, reporter info, timestamps
console logs captured with each report
network requests, including status codes and flagged failures
environment data: browser, OS, device, viewport, page URL
screenshots and session replay context
the tasks the reports are organized into
That means questions like these work as plain prompts:
"Show me the most recent bug reports in bugdump and summarize the top issue."
"Are there any bugdump reports where users hit a 500 on /api/orders? Summarize the network requests."
"Find the bugdump task about the checkout flow and show me the console errors from its reports."
The assistant reads the raw telemetry and answers with a synthesis — and because tools like Claude Code also have your codebase open, the next step is often "now find where that request is made and fix the null check." The bug report and the fix happen in the same session.
Setting it up
Updated October 2026: Claude, Claude Code, ChatGPT and VS Code can now connect by signing in to BugDump, with no token — see Connecting Claude, ChatGPT or VS Code to BugDump. MCP access is free on every plan.
Cursor and other clients connect with a personal token, one per person:
Open your project's MCP server page and create a token. Any member can create one, and it acts as you — the assistant sees what you can see. Leave Allow changes off for read-only access, or check it to let the assistant move tasks, comment, set priority and assign.
Copy the
claude mcp addcommand the page shows, or commit the shared.mcp.jsonit offers. That file reads the token from aBUGDUMP_MCP_TOKENenvironment variable, so no secret lands in your repository.Restart or reload your MCP client. The BugDump tools appear alongside your other MCP servers.
If a token leaks, revoke it on the same page and it stops working immediately. When a contractor rolls off the project, removing them from the team revokes all of their tokens at once.
The MCP server is included on every plan, Free too. Why we moved from a shared key to personal tokens is in MCP Server Authentication: Why AI Agents Need Personal Tokens.
A realistic workflow
Here is how this plays out on an actual bug:
A user reports "checkout button does nothing" through the widget on your site. The report arrives with console logs, network activity, and a replay.
In Claude Code, you ask: "Look at the newest bugdump report on the checkout flow. What's failing?"
The assistant fetches the report, finds an uncaught
TypeErrorin the console, and notices the/api/cartrequest returned a 500 just before it. It summarizes: the client assumed the cart response always contains ashippingobject, and the API returned an error page instead.You ask it to find the code path. Since Claude Code has your repository, it locates the cart handler, adds the missing error handling, and you review the diff.
You ask the assistant to move the task to Done and comment with a summary of the fix — the change shows up under your name, marked "via MCP" — and, if you use integrations, close the linked GitHub issue.
No copy-pasting logs, no reconstructing the failure from a screenshot, and no back-and-forth with the user — because the report captured everything at the moment it happened.
What this doesn't replace
An MCP-connected assistant is a fast reader and a good hypothesis generator; it is not a substitute for judgment. You still decide which bugs matter, whether a proposed fix is correct, and whether the failure points at a deeper design problem. Treat the assistant's summary as a head start on triage, not a verdict.
It also can't conjure context that was never captured. The quality of AI-assisted debugging is bounded by the quality of the reports — which is the argument for capturing console, network, and replay data automatically rather than hoping users describe them.
The bigger picture
Bug workflows have always had a translation layer: users describe symptoms, developers reconstruct causes. Automatic capture removed the first half of that translation — the report carries the technical facts. MCP removes much of the second half — the assistant reads those facts and maps them to code.
What remains for you is the interesting part: deciding what to build and reviewing the fix. If your bug reports currently live in a tool your AI assistant can't see, that's the gap worth closing first. With BugDump, closing it is a personal token and one command.