MCP server for detecting and downloading media content (videos, audio, images) from web pages, with support for HLS streams
This server defines 2 tools with complete input/output schemas in a JSON manifest (v3/mcp/mcp.json). Both tools have descriptions and reasonably detailed parameter documentation. However, the tool naming deviates from verb-noun convention (find_media, download_media use underscores but the internal command enums use hyphens like 'find-media'), and parameter descriptions, while present, lack some specificity around constraints and error cases. The outputSchema documentation is well-structured with oneOf unions for success/error cases, but the server lacks critical safety features like idempotent hints, readonly annotations, and comprehensive error guidance. The parameter 'cmd' is required and enum-constrained, which is good, but the redundancy of requiring a 'cmd' field in both input and output responses adds friction. Overall, definitions are above-average completeness but lack LLM-optimization and safety annotations.
Queue media files for download. Opens the downloader interface with the specified media items already added. Supports HLS streams.
Detect media content (videos, audio, images) from web pages. Can scan existing tabs or open new tabs and monitor for media in real-time.
Tool naming uses underscores (find_media) but internal command enums use hyphens (find-media). This mismatch creates cognitive friction and could confuse agents about the canonical tool name.
Missing toolAnnotations. Neither tool declares readOnlyHint, destructiveHint, or idempotentHint. find_media should be marked readOnlyHint=true (risk: READ_ONLY); download_media should be marked destructiveHint=true (risk: WRITE). These hints guide agents on retry safety and side-effect awareness.
Error handling is minimal. Error responses include only 'type' (optional) and 'message' fields. No guidance on recovery steps, error categorization (retryable vs. user-fixable), or what the LLM should try next. E.g., if find_media fails because a tab is inaccessible, the error should suggest 'Try opening a new tab with the query parameter.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2025-06-18+ | v2 |
Parameter descriptions lack format and constraint details. E.g., 'tabId' is described as an integer but no bounds are given (valid range?). 'query' accepts 'chrome.tabs.query argument' but this is opaque to an LLM unfamiliar with Chrome APIs. 'jobs' array in download_media should specify min/max counts and URL format constraints (e.g., must be valid HTTP/HTTPS URLs).
Output schema for find_media uses a custom 'cmd' enum field (media-detected) in the success case, mirroring the input 'cmd' requirement. This is non-standard and forces agents to parse a command field rather than relying on tool response structure. Standard MCP responses use object types with domain-relevant fields.
No pagination support for potentially large result sets. If find_media detects many media items on a page, the response array could be huge, wasting tokens and blowing context windows. The tool should support limit/offset or cursor pagination and state this in the description.