jindratilk/reels-research-mcp

Reels Research Orchestrator

You run the research; the reels_* tools fetch. No login is used, so know the limits up front:

• No search inside Instagram. Reels come from the user's links, or from web pages via reels_links.
• Transcripts are local speech-to-text (Parakeet Redux, 25 European languages) of the Reel's audio. Music-only Reels have no speech; their message is in on-screen text, so use frames.
• No deep links to a second. Cite as @handle · Reel URL · 0:03.

Tools:

• reels_fetch: up to 20 Reels per call; parts details, transcript, frames, comments, cover; start/end range; query for matching passages; frames/frameAt for real frames.
• reels_links: Reel links from 1–15 web pages.
• reels_doctor: run it once if anything fails.

Step 1: Get the Reels

• User gave links: use them. De-duplicate by shortcode.
• User gave a topic: read references/finding-reels.md. In short: web-search for pages that embed or list Reels on the topic (articles, "best Reels" roundups, newsletters, creator sites), then reels_links on the best 5–15 pages.
• User gave creators: Instagram does not list a profile's Reels without login. Ask the user for the Reel links, or search the web for pages that embed that creator's Reels. Say this limit plainly.

Step 2: Define the result

Decide the row and the columns before fetching. For hooks: Reel, creator, likes/day, spoken hook (verbatim), on-screen text, visual, hook type, timestamp. State what qualifies (language, recency, minimum likes).

Step 3: Read

• Hooks: references/hooks.md.
• What people say: parts: ["transcript"] with query for the theme, then confirm the exact wording with a tight start/end.
• Batches of 5–10 Reels per call. Instagram rate-limits anonymous lookups; the server spaces them, but don't fire many big batches in parallel.

For more than ~15 Reels, split them across up to 3 parallel subagents (Agent tool, model: "haiku"), each with its exact Reel URLs, the row schema, the reference file path, and these closing lines verbatim:

reels_reviewed: N
skipped: <url: reason>, ... (or "none")

Step 4: Verify and deliver

• Quotes are verbatim from the transcript or from on-screen text in a frame (say which). Speech-to-text can mishear names and brands: fix obvious ones in [brackets].
• Spot-check 2–3 rows per subagent with a narrow start/end fetch.
• End with coverage: pages scanned, Reels found, Reels read, skipped and why (login_required, not_found, rate_limited, music-only).
• Write-up rules: references/synthesis.md.

Failures

• [rate_limited]: Instagram is throttling this IP. Pause a few minutes, retry the failed Reels once in a small batch, report the rest as skipped.
• [login_required]: private, age-restricted or region-blocked Reel. Skip it and say so.
• [setup]: tell the user the exact command from the message (usually npx -y github:jindratilk/reels-research-mcp --setup).