Here is a workflow that was normal until recently.
A bug report arrives. You open the dashboard, read it, find the console error, copy it. Switch to your editor, paste it into Claude Code, add some context. The agent asks what the request returned. Back to the dashboard, find the network log, copy the response, paste it. And so on.
You are a clipboard. The agent is doing the reasoning and you are doing data transport between two systems that could talk to each other.
So we made them talk to each other.
The MCP server
BugDump ships a Model Context Protocol server. Connect it to Claude Code and the agent reads your bug reports directly. Seven tools:
Tool | Returns |
|---|---|
| The project this key is bound to |
| Tasks, for finding one |
| One task, by public ID or task ID |
| A task's history |
| The reports attached to a task |
| One report — console, network, screenshots |
| Complete telemetry, paged |
The whole workflow collapses to one sentence:
Look at task #42 in BugDump and fix it.
The agent reads the task, pulls the attached reports, correlates the console error with the failing request, finds the matching code in your repo, and proposes a fix. You never open the dashboard.
Why this works better than pasting
The obvious benefit is that you stop being a clipboard. There are two less obvious ones.
The agent asks better questions than you do. When you copy a stack trace, you have already decided what is relevant — and you decided before understanding the bug. An agent with direct access can follow the evidence: read the console, notice a failing request, page through the network log for its full body, check whether other reports on the same task show the same failure. That is a lot of small lookups. Each is trivial to serve and tedious to perform by hand, which is precisely why they don't get done.
It can compare across reports. Ask which browsers are affected by task #17 and the agent reads every attached report and tells you. That is genuinely tedious work — the sort a human triager skips, and the sort that identifies a bug as browser-specific in ten seconds instead of an afternoon.
The design decisions
Two choices shaped how this works.
Keys are permanently bound to one project. There is no project parameter on any tool. An MCP key is scoped at creation and cannot be widened afterward. When you hand a key to an autonomous agent, "this key can only ever see one project" is a much stronger guarantee than "the agent has been told to only look at one project."
MCP keys are also read-only and separate from widget keys. The key in your frontend cannot read reports; the key that reads reports never goes near a browser.
Summary by default, full telemetry on request. get_report returns the last 50 console entries with network bodies stripped. A report can contain thousands of entries, and dumping all of it into a context window on every lookup is how you end up with an agent that has read everything and understood nothing.
But truncation that hides itself is worse than no truncation. So the summary states what was trimmed, and get_report_telemetry pages through the complete history — console, network with bodies, actions — in chronological chunks until nothing remains.
The rule that matters: a summary showing nothing is not evidence that there is nothing.
Prompts that work
Look at task #42 in BugDump and fix it.
What do the open tasks in BugDump have in common?
Read the reports on #17 and tell me which browsers are affected.
Check task #8 — is this the same bug as #3?
That last one is our favorite. Deciding whether two reports are the same bug requires comparing two sets of telemetry in detail, which is exactly the work everyone skips and exactly the work that keeps duplicate tasks out of your board.
The larger point
We built the capture layer first because incomplete bug reports were the problem. The MCP server was not the plan; it became obvious afterward.
A complete report — console, network, actions, replay, environment — turns out to be a well-formed input for an agent. And an agent cannot ask a follow-up question. It cannot email a user to ask which browser they used. It gets one shot at the evidence.
Which means the discipline that makes a report good for a human is the same discipline that makes it usable by a machine: capture everything, at the moment it happens, without asking.
Getting that right had a payoff we did not design for.
Update, October 2026: MCP access is now included on every plan, Free too, and Claude, ChatGPT and VS Code can connect by signing in — see Connecting Claude, ChatGPT or VS Code to BugDump.