Cursor
Cursor
Settings → MCP, or .cursor/mcp.json
Read the bug queue with Cursor, Claude, Codex, and Gemini.
Connect the BugTrap MCP and ask about the bugs already filed. The agent lists the queue and reads a report. The screenshot, the replay, and the ticket stay in the dashboard, where you can still assign, comment, and close.
Clients
Same URL. The config depends on the client.
Claude Code
Claude
Project .mcp.json
Codex
OpenAI
~/.codex/config.toml
GitHub Copilot
VS Code
.vscode/mcp.json
Gemini CLI
Gemini
~/.gemini/settings.json
Devin Desktop
Devin
~/.config/devin/mcp_config.json
Workflows
Ask about a bug that is already filed.
The server exposes two tools. The agent calls list_issues to find a bug, then get_issue with that bug's id.
list_issues
Triage the queue
Ask what's open, what's critical, or what matches a search. Each row includes the id, the BT- code, the title, the assignee, and the page it was filed on.
get_issue
Read one report
Pass the id from that list. The report includes the description, the page URL, the environment, a short console excerpt, failed requests, the HTML and CSS of a picked element when one was captured, the ten most recent comments, and a link back to the bug.
get_issue
Fix it in the same chat
A coding agent can take that report and edit the repo. The BugTrap tools still only read. Closing, assigning, and commenting stay in the dashboard.
get_issue
Catch up on the thread
The ten most recent comments come back with the report, public and internal. The agent can summarize them. Posting a comment stays in the dashboard.
The report
What the agent gets, and what stays here.
A report is text. It is written for an agent that needs the bug, the page, and the errors, then a link when it needs the replay.
The whole queue
One question can cover the open bugs. Filter by status and priority, or search the title and description. The list is capped at 50.
The errors, shortened
Console lines come back with errors first. Failed requests include the method and path. BugTrap tokens, bearer tokens, and secret-looking fields in those lines are redacted.
The picture stays on the bug
The screenshot, the annotation, and the session replay stay attached to the bug. The report names them and ends with the dashboard link.
The ticket stays editable here
The tools only read. You still assign, comment, and close in the dashboard. Revoke the token in Settings → MCP and that client loses access.
Connect
A token, a config, then a question.
The token belongs to one person and one workspace. Revoke it from Settings → MCP and that client loses access.
- 01
Create a token
In the dashboard, open Settings → MCP. Admins and members can each keep ten tokens. The secret is shown once, to the person who created it.
- 02
Paste the config
Copy the config for your client and replace YOUR_TOKEN. The workspace API key on the install snippet is for the widget, not this connection.
- 03
Ask about a bug
Ask the agent to list open bugs, or to read one by the id from that list. The BT- code is a label. The tool uses the id.
FAQ
What the agent can see.
The client, the token, and what leaves the workspace with the report.
What is an MCP? What is the BugTrap MCP?
MCP (Model Context Protocol) is an open way for an agent to call tools outside the chat. The BugTrap connector gives that agent two tools for one workspace: list bugs, and read one report. Describe the bug you want. The agent looks it up.
Which agents can I connect?
Any client that can call a remote MCP server and send a bearer token. Configs on this page cover Cursor, Claude Code, Codex, VS Code with GitHub Copilot, Gemini CLI, and Devin Desktop. ChatGPT and the Claude app add custom connectors with OAuth, which this server does not use. For those models, use Codex or Claude Code.
Do I need to know how to code?
No. Create a token, paste the config, and ask a question. Reading the queue is a conversation. When the agent is inside an editor, it can also use the report to change the code.
What can the agent see?
The list includes the id, the BT- code, the title, the status, the priority, the type, the assignee, and the page. The report adds the description, the page URL, the environment, a short console excerpt, failed network requests, the HTML and CSS of a picked element when one was captured, the ten most recent comments, and a link to the bug. Public and internal comments are both included.
Can the agent see the replay?
The report names the screenshot, the annotation, and the replay when they are attached, and it includes the dashboard link. Those files stay on the bug page.
Can the agent change a bug?
The tools only read bugs. Closing, assigning, and commenting stay in the dashboard.
Can the whole team use it?
Yes. An admin or a member of a workspace with an active subscription can create a token. Each person can keep ten. An admin can see every token in the workspace and revoke one. A member sees the tokens they created. Bugs the agent reads are the bugs already in that shared workspace.
What is sent to the model?
The report text is sent to the model that client uses. Console lines are shortened, and BugTrap tokens and bearer tokens in those lines are replaced. When a reporter picked an element, its HTML and CSS are included and shortened the same way. Fields whose names look like passwords, secrets, or API keys are redacted. The screenshot, the annotation, and the replay are not sent.
How is this different from the dashboard?
The dashboard is where the team filters, assigns, comments, and closes. MCP lets an outside agent read that same queue as part of the chat it is already in, including a coding agent that then edits the repo. The widget still files the report.
How do I get started?
Open Settings → MCP and create a token. Copy the config for your client from this page, replace YOUR_TOKEN, and ask the agent to list open bugs. The widget API key does not work here. The secret is shown once.
BugTrap
The next bug is one question away.
The widget still files the report, with the screenshot, the markup, the element, and the replay. MCP is how your agent reads the queue. See every feature.