Connect the public, key-free "AI video intel radar" MCP to Claude Code / Codex and ask which AI-video formats emerged today, where a trend started, or what a model shipped this week
Chinese creator @GaryAIytb (Gary Cao) runs an AI intel radar that scans public posts on Bilibili, YouTube, Douyin, TikTok, Xiaohongshu, X, GitHub and more every morning to flag AI-video formats that are just taking off. He had Claude build a remote MCP server; one command connects it to Claude Code / Codex, then you ask in Chinese: what's new or heating up today, when did this format start and is it too late, what did Kling ship this week. We called all 7 tools directly over the MCP protocol (no client)—anonymous and returning real data—and ran the open-source version's free collectors locally: 1 min 46 s, zero cost.
Visit site AI 情报站 MCP (lab.ytdeepsea.com) + Claude Code / Codex ↗- Pricing
- Free
- Region
- China
- Where it's from
- Chinese community
- Added
- Oct 28, 2026
- Tested on
- Oct 28, 2026
- Version we used
- Public MCP ai-radar 1.0.0 (protocol 2025-06-18), issue 2026-10-28 #6; intel-radar commit 0aff8a3 (2026-10-24)
- Original author
- Gary曹淦 @GaryAIytb ↗
- Our test
- Partly tested
- Cost / time
- The public MCP and site are free, anonymous and keyless. Self-hosting: the README puts one issue at ~110 TikHub calls ≈ $0.20 plus a few jiao of DeepSeek; free sources only costs nothing.
The task
AI-video creators need daily topic ideas, a read on whether a format is fresh or saturated, and model-update tracking—without scrolling several platforms every day or having an agent scrape pages and burn tokens.
Steps
- Claude Code: `claude mcp add --transport http ai-radar https://lab.ytdeepsea.com/mcp`; Codex: `codex mcp add ai-radar --url https://lab.ytdeepsea.com/mcp`. Start a new session afterwards. Other clients are covered at lab.ytdeepsea.com/radar/agent/.
- No need to remember tool names—just ask. The server exposes 7 read-only tools: radar_today, radar_video_topics (filter by new / rising / steady / watch / fading and by platform), radar_x_topics, radar_tools (repos / models / news), radar_topic_origin (trace a format back), radar_search (past issues), radar_list_issues.
- Each morning: which AI-video formats are new or rising today, one line plus a link each.
- For a format that's climbing, ask for its origin: earliest post, platform, date, and how far along it is—i.e. is it too late to jump in.
- Track a tool or model: "what has Kling done this week" returns dated headlines, X buzz and news items.
- Wire it into your routine (author's ideas): every morning match new/rising formats against your niche and write 2–3 topic cards; give a team the day's headlines as stand-up notes; check reference videos and reasoning before scripting.
- Assistants without MCP: give them https://lab.ytdeepsea.com/radar/agent.md and they'll read latest.md, a dated .md or the JSON.
- To track another niche, self-host the open-source garycaogan/intel-radar (Python stdlib only). Try a free run with `python3 -m radar run --only bili_search,bili_popular,bili_hot,github_trending,hn,hf,news --skip-llm`; a full issue needs TikHub and an OpenAI-compatible model key.
Prompt
Use ai-radar: which AI-video formats are new or rising today? Give one line, the numbers and the radar link for each. Then pick one rising format, trace its origin, and tell me where it first appeared and whether it's too late to start. Answer only from what the tools return; say so if something isn't there.
What the author says
@GaryAIytb: Claude wrote the MCP. Asking Codex "which AI-video formats are new or rising today" returned 1 new and 2 rising (e.g. "Claude hacks smart TVs", already followed by 4 accounts). Tracing "Hindi vegetable-family animation" found the earliest TikTok on June 30, YouTube copies from August, a ~7k-follower account hitting 1.67M views, and the verdict "not too late, but the window is closing." A new issue lands around 6 a.m. Beijing time; numbers are computed by code from public data, answers cite sources and admit gaps. His previous radar post had 20k views.
Our test
We did not connect it from a Claude Code / Codex client (neither is installed on our box); we called it with curl over MCP Streamable HTTP: initialize → tools/list → tools/call, exercising radar_list_issues, radar_video_topics (new), radar_topic_origin (vegetable family) and radar_search (Kling). We also shallow-cloned garycaogan/intel-radar (commit 0aff8a3, 2026-10-24) and ran the README's free trial command on Python 3.13.
What we found when we tried it:initialize works with no key (server ai-radar 1.0.0); tools/list returns 7 tools, all flagged readOnlyHint. The issue list showed 5 issues for Oct 24–28 (Oct 28 is issue 6). Oct 28 had exactly one "new" format: celebrity-character AI song-and-dance MVs (Bilibili; 6 accounts, 5.663M views total, a 10k-follower account at 1.572M on one video), with reasoning and reference links. The vegetable-family trace matched the article: earliest TikTok on 2026-06-30, a sibling hit at 1.678M views. Searching Kling returned 16 dated items over 7 issues. The open-source free run collected 825 items in 106 s (641 from Bilibili) with zero TikHub/model spend—but the repo has no MCP server code.
How to get in
The original is a Chinese X Article (login may be required). The MCP, site and agent.md are public with no login or key. Content and answers are in Chinese by default—ask your agent to reply in English; the reference links (Bilibili, TikTok, YouTube) open as usual.
Who it suits
AI-video, short-drama and cross-border content creators and small teams already using Claude Code / Codex who want an agent to handle daily topic scouting.
Pitfalls
- One issue per day (around 6 a.m. Beijing); repeated questions on the same day hit the same issue. Origin tracing only covers new and rising formats.
- "New / rising" follow the author's own rules (two independent signals, etc.) and lean toward AI video and Chinese platforms—not whole-web popularity.
- "Earliest found" is the earliest post platform search returns; the real origin may be older (the tool says so itself).
- Video titles and posts are third-party content; the server instructions tell the model to treat them as data, never instructions—hold your own agent to that too.
- In our check: the article says the open-source version ships the same MCP, but as of 2026-10-28 the repo's latest commit (2026-10-24) contains no MCP server code—confirm the repo has been updated before planning a self-hosted MCP.
- It's a personally run public service; uptime and data rules can change. Click through to the source videos before acting on anything important.
More in this category
- Use a web app once and get an AI write-up of how it worksResearchREAClaude CodeChromeFree tools; needs Claude or Codex◐ Partly testedOverseas
- 1,050 memory recalls audited: what makes agent memory pay offCodingNowledge MemCodexClaude CodeFree tier: 50 memories; review uses model quota○ Not tested by usChina
- Give your agent 300 search sourcesResearchArgoClaude CodeCodexFree; works without API keys◐ Partly testedChina
- Align terms, check facts, then decideCodingsinfire-skillsCodexgrill-meSkills free; needs Codex or Claude◐ Partly testedChina
- 1,169 dossiers + two evidence-based growth research methodsResearchBeRichClaude CodeIndie hackersSkill free; full paths locked in Pro◐ Partly testedChina