Feedbug MCP server: read visual bug reports (screenshot, console, DOM context) and let your AI agent trace them to source and fix them.
Feedbug MCP demonstrates solid definition quality with well-structured tool names, comprehensive descriptions, and proper schema definitions. All 7 tools follow verb_noun naming (feedbug_list_bugs, feedbug_get_bug, etc.). Descriptions are detailed (150-350 chars) and include WHEN/DO NOT USE guidance, matching the pattern:tool-description baseline. Input schemas use Zod with proper types and constraints. However, tool #7 (feedbug_resolve_bug) has a minimal description (7 words) that falls below the 20-char threshold, and output schemas are not explicitly documented in the code. Error handling descriptions are present but lack actionable recovery guidance (e.g., 'Bug not found' does not suggest next steps). Tool composition is clean, each tool has one responsibility, and parameter naming is consistent (id, status, page_url, limit). Annotations are properly declared (readOnlyHint, destructiveHint, idempotentHint, openWorldHint).
Post a comment on the tracker issue linked to a bug. Returns: a confirmation message once the tracker accepted the comment. Use when: you want to document a fix, ask the reporter for a reproduction, or leave an update the whole team can read on the issue. Do not use when: the bug is tracked in GitHub — only Linear-routed bugs accept comments today and the call errors otherwise. Never use it to talk to the user running you; answer them directly instead. Behavior: writes to Linear and is NOT idempotent — every call adds another comment, and this server can neither edit nor delete one. The comment is public to everyone with access to the issue and is authored by the Linear account that installed Feedbug, so send a single finished comment rather than successive drafts.
Read one bug report in full. Returns: a JSON object with the title, the Feedbug and tracker statuses, the tracker issue URL, the page URL, the screenshot URL, the session replay URL, the viewport, the user agent, the click coordinates as percentages, and the captured DOM context. Use when: you picked a bug out of feedbug_list_bugs and are about to reproduce or fix it. Do not use when: you only need the DOM tree (feedbug_get_html_context), only the browser environment (feedbug_get_diagnostics) or only the replay link (feedbug_get_replay). Behavior: read-only and idempotent, scoped by FEEDBUG_PROJECT_KEY to a single project. An unknown id, or one belonging to another project, returns a "Bug not found" error.
Read the browser environment a bug was reported from. Returns: a JSON object with the page URL, the viewport size, the user agent and the name of the reporter. Use when: you need to reproduce a bug under the same browser, screen size or account, or to judge whether a report is device-specific. Do not use when: you also want the screenshot, the DOM or the tracker link — this is a strict subset of feedbug_get_bug, so one call there replaces this one. Behavior: read-only and idempotent, scoped by FEEDBUG_PROJECT_KEY to a single project. Fields the widget could not capture come back null; an unknown id returns a "Bug not found" error.
feedbug_resolve_bug has minimal description (7 words: 'Mark a bug as resolved in Feedbug.'). Descriptions under 20 chars cannot adequately explain WHEN to use the tool or what it returns. LLMs cannot distinguish this from similar state-change tools without richer context.
Output schemas are not explicitly documented in code. While responses are JSON-serialized, the structure of returned fields (e.g., what feedbug_list_bugs returns per bug, what feedbug_get_bug includes) is only described in natural language. LLMs benefit from formal schema definitions to plan downstream calls.
Error messages lack actionable recovery guidance. Descriptions mention errors (e.g., 'Bug not found', 'No HTML context captured') but do not suggest what the LLM should do next. Per pattern:recovery-guide, errors should include hints like 'Try feedbug_list_bugs() to find a valid bug ID.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2026-07-28+ | v2 |
Read the HTML and DOM context captured at the instant the tester clicked on the bug. Returns: a JSON object with the CSS selector path, the framework component path (Vue, React, Angular, Svelte), the clicked element itself (tag, id, classes, text, attributes), its ancestors and its siblings. Use when: you need to map a report onto a source file — this is the shortest path from a bug to the component that renders it. Do not use when: the bug is about a behaviour rather than one element; feedbug_get_replay shows what the tester actually did. Behavior: read-only and idempotent, scoped by FEEDBUG_PROJECT_KEY to a single project. A report captured without DOM context errors with "No HTML context captured for this bug" rather than returning an empty tree. Component paths survive a Vue build but are often absent from minified React, in which case fall back to the selector path.
Read the session replay URL of a bug, an rrweb recording of the last 30 seconds of activity before the report was sent. Returns: a one-line message holding the URL to open in a browser. The recording itself is never inlined, so it costs no context. Use when: the written report is not enough to reproduce the bug and you want to hand the user a link to watch it happen. Do not use when: you already called feedbug_get_bug, which returns the same URL alongside the rest of the record. Behavior: read-only and idempotent, scoped by FEEDBUG_PROJECT_KEY to a single project. A bug with no recording returns "No session replay available" as an ordinary result rather than an error, so the call is always safe to attempt.
List the bug reports captured by the Feedbug widget, newest first. Returns: a JSON array carrying, per bug, its UUID, tracker identifier (e.g. PRE-13), title, status, page URL, framework component path, session replay URL and creation date. Use when: you are looking for a bug to work on, checking what is still open, or reviewing every report filed on one page. Do not use when: you already hold a bug id — feedbug_get_bug returns that one report with its screenshot and DOM context. Behavior: read-only and idempotent. The server authenticates with the FEEDBUG_PROJECT_KEY environment variable and can only ever see that single project. There is no pagination: the API truncates to limit, so narrow the set with status or page_url instead of paging. No match returns an empty array, not an error.
Mark a bug as resolved in Feedbug.
feedbug_add_comment is a WRITE tool but lacks a dry-run or confirmation step. Per pattern:confirmation-request, irreversible operations (especially public comments on shared issues) should support a preview or confirmation before execution to prevent accidental posts.
feedbug_resolve_bug lacks input validation description. The 'id' parameter accepts UUID or tracker identifier (case-insensitive), but the description does not specify the format or provide examples. LLMs may pass invalid IDs without clear guidance on what constitutes a valid ID.