@bugzar/mcp

Local-first MCP server for Bugzar reports. Point an AI coding agent (Claude Code, Cursor, …) at a captured self-contained report HTML file and it can query the repro steps, console errors, failed requests, and the same curated AI context "Copy for AI" produces — directly, no copy/paste.

No cloud account. No network call unless you pass a URL. The report HTML already contains everything (it's the same file you'd double-click to replay offline) — this server just reads it locally and answers structured queries instead of you pasting one static Markdown dump.

Install

claude mcp add bugzar -- npx -y @bugzar/mcp

Or manually in your MCP client's config:

{
  "mcpServers": {
    "bugzar": {
      "command": "npx",
      "args": ["-y", "@bugzar/mcp"]
    }
  }
}

Usage

In Claude Code / Cursor, once connected:

Open the report at ./bug-report.html and tell me why the checkout request failed.

The agent will call open_report (local path by default), then get_failed_requests / get_console_errors / get_ai_context as needed — same data Copy for AI curates, queryable instead of a single static block.

Tools

| Tool | Purpose | |---|---| | open_report | Open a report — local path by default, or url (explicit opt-in only). Returns a handle + diagnostic summary. | | get_repro_steps | The same auto-synthesized reproduction steps shown in the Steps panel. | | get_console_errors | Captured console errors + top stack frames, capped by limit. | | get_failed_requests | Failed network requests (status ≥ 400 or transport error), capped by limit. | | get_network_request | Full detail (headers + body, already redacted) for one request by index. | | get_system_info | Browser/OS/viewport/connection + Web Vitals. | | get_ai_context | The exact Markdown Copy for AI produces — symptom, failing request, errors, repro, where-to-look. |

Every tool other than open_report takes the handle returned by open_report. Reports opened in one process are not shared across connections — open the report again if your client restarts the server.

Privacy

Values in a report are already redacted at capture time (Authorization/Cookie/ JWT/PII) — this server doesn't add or bypass any of that; it only reads what's already in the file. Passing url to open_report is the only network call this server ever makes, and only when you explicitly ask for it.