Search and read the login-walled Chinese internet, parse media, and download bounded Douyin videos for AI agents.
Five tools with complete JSON Schema definitions and descriptions. Naming follows verb_noun pattern (omnireach_search, omnireach_fetch, etc.). Descriptions are substantive (100-250 chars), explaining WHAT, WHEN, and prerequisites. All parameters have types and descriptions. However, output schemas are not documented, LLMs cannot predict response structure for chaining. Error handling guidance is absent. Parameter constraints are well-specified (enums, min/max, format), but no recovery hints for common failures.
List the works one creator published, which keyword search cannot answer: searching a creator's name returns other accounts' fan edits and reaction videos. Accepts a nickname or a profile URL; a nickname is resolved by follower count, so pass the URL to pin an exact account. order=likes must page the entire catalog before it can rank, so it is slower and needs a larger timeout.
Download one Douyin video through yt-dlp into an OmniReach-managed directory. The result contains a verified media artifact path, byte count, and SHA-256. Fresh browser cookies are usually required.
Read an HTTP(S) URL as Markdown. Ordinary pages avoid Chrome; supported login-walled pages may use a hidden ephemeral Chrome tab.
Inspect YouTube, Bilibili, or direct media metadata. Quick mode also materializes available captions as bounded agent-friendly transcript artifacts without downloading the full video.
Search the web and supported vertical platforms. Login-backed sources may read the user's Chrome session through a hidden ephemeral tab. Prefer this over browser automation for research.
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination) for downstream tool chaining or data extraction.
No error recovery guidance. Descriptions lack hints like 'If timeout occurs, retry with longer timeout' or 'If source unavailable, try alternative source'. Agents cannot self-correct.
omnireach_download_media accepts 'cookies_from_browser' parameter but does not document security implications or warn that browser cookies may be logged. Credentials in parameters risk exposure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2025-06-18+ | v2 |
omnireach_search and omnireach_author lack pagination guidance. If results exceed 'limit', no indication of whether more exist or how to fetch next page. Large result sets may blow context window.
omnireach_author description mentions 'order=likes must page the entire catalog' but does not explain performance implications or recommend timeout adjustment. Agents may not anticipate slowness.