A keyless MCP server and research skill for Instagram Reels. It hears what's said, sees what's on screen, and breaks down how every Reel opens, on your Mac.
MIT license · Local speech-to-text (Parakeet Redux, 25 languages) · Claude Code & Codex plugin · Any MCP client
[0:00] I know for a fact you're wasting half your day…
[0:00] I do a $106 million launch, and it starts with…
[0:01] If you're starting a business this year, watch…
0:00
0:01
0:02
0:00
“I know for a fact you're wasting half your day on tasks that belong in the trash.”
“I do a $106 million launch, and it starts with one sheet of paper.”
“In 1989 I was the kid behind the counter of a hobby shop.”
“If you're starting a business this year, watch this before you spend a single dollar.”
Illustrative session with AI-generated frames. Tool names, parameters and output shape are real.
WORKS WITH ANY MCP CLIENT
How it works
Reels carry their message in three channels at once: speech, on-screen text and the picture. The agent gets all three.
Instagram has no anonymous search, so the web is the index. The agent searches the web for articles, roundups and newsletters on your topic, and reels_links pulls every embedded or linked Reel out of them in one call. Or just paste your own links.
[0:00]So there is no perfect length for a reel, but generally we recommend reels under three minutes.
[0:11]One, think about what your hook is.
[0:13]How are you going to get someone's attention in the first three seconds?
Instagram doesn't share its captions without a login, so the server listens itself. It fetches only the Reel's audio and transcribes it on your Mac with Parakeet Redux: 178 MB, 25 European languages, timestamps per sentence. Nothing is sent to a cloud.
0:00
0:01
0:02
0:03Real frames, vertical, at any second. Burned-in subtitles, text overlays, b-roll and cuts become readable for the agent. ffmpeg reads only the bytes around each frame.
| Spoken hook | On screen | Type |
|---|---|---|
| “This setup saved me 20 hours last week.” | “20 HOURS” | Result first |
| (no speech, trending sound) | “POV: your first client says yes” | Trend |
| “Small bakery owners, listen.” | none | Niche callout |
Coverage: 9 pages scanned · 23 Reels found · 18 read · 3 music-only · 2 login_required
Every row says where it came from: the transcript or a frame, cited as @handle, Reel URL and second, and every answer ends with what was found, read and skipped.
Private by design
Public Reel data in, a cited answer out. Speech-to-text runs locally on Apple silicon or any CPU; no account, no API key, no cloud transcription.
Only what Instagram shows any logged-out visitor. Your account is never involved, so it can't be flagged.
Transcript for what's said, real frames for what's written and shown. Music-only Reels are reported as such, not guessed.
Verbatim quotes with their source and second, a coverage line, and spot-checks by the skill.
The tools
Read up to 20 Reels: exactly the slice and the channels you need.
{ "reels": ["instagram.com/reel/…"],
"parts": ["transcript", "frames"],
"start": 0, "end": 5, "frames": 3 }
Turn web research into Reels: every Reel embedded or linked on up to 15 pages.
{ "pages": ["https://blog…/best-reels",
"https://news…/story"] }
Checks every piece on your machine and says how to fix what's missing.
✓ Reel access: @mosseri, audio yes, video yes ✓ speech-to-text: Parakeet Redux loaded
Use cases
Spoken hook, on-screen text and visual of every Reel's first seconds, typed by pattern.
“Break down how these 12 Reels open and which pattern gets the most likes per day.”
Web search plus reels_links turns a topic into a list of Reels to study.
“Find Reels about meal-prep for busy parents and summarise the angles they take.”
Verbatim quotes with the second they were said, from any of 25 European languages.
“What do these coaches say about pricing? Quote them.”
Frames show framing, text overlays, subtitles and cut pace across a set of Reels.
“How do these brands use text on screen in the first two seconds?”
Paste a competitor's Reel links and get a structured comparison of message and format.
“Compare these 8 Reels from two competitors: topics, hooks, calls to action.”
Music-only Reels are flagged, so trend-and-text formats separate cleanly from talking ones.
“Which of these Reels rely on a trending sound instead of speech?”
Why it's different
| Typical Instagram scraper | Reels Research MCP | |
|---|---|---|
| Access | Your login or session cookies | No login, no cookies: only public data |
| Speech | Cloud transcription, or none | On-device Parakeet Redux, 25 languages, timestamps |
| On-screen text | Not covered | Real vertical frames at any second |
| Finding Reels | Logged-in search | Your links, or Reels pulled from web pages |
| Honesty | Silent gaps | Coverage line and a reason for every skipped Reel |
| Method | Left to you | Research skill: hooks taxonomy, subagents, spot-checks |
Install
Paste one prompt into your agent. It installs the server and the skill, sets up on-device speech-to-text, and checks that everything works.
Install Reels Research MCP for me from https://github.com/jindratilk/reels-research-mcp. Follow the “Install instructions for AI agents” in its README for the client you are running in: the plugin (MCP server + research skill) in Claude Code or Codex, otherwise the MCP server config. Run its --setup step for local speech-to-text, then run the doctor and tell me whether everything works, and if I need to restart anything.
$ claude plugin marketplace add jindratilk/reels-research-mcp $ claude plugin install reels-research@reels-research $ npx -y github:jindratilk/reels-research-mcp --setup
$ codex plugin marketplace add jindratilk/reels-research-mcp $ codex plugin add reels-research@reels-research $ npx -y github:jindratilk/reels-research-mcp --setup
{
"mcpServers": {
"reels": {
"command": "npx",
"args": ["-y", "github:jindratilk/reels-research-mcp"]
}
}
}Honest limits
Instagram offers logged-out visitors no search, hashtag pages or profile listings. Reels come from your links or from web pages via reels_links.
Instagram's own auto-captions are only served to logged-in users, so transcripts come from on-device speech-to-text. It can mishear names; noisy audio lowers accuracy.
View counts are hidden from logged-out visitors. The server reports likes, likes per day and comment counts.
Instagram throttles anonymous lookups; the server spaces them and says when to wait. Private, age-restricted or blocked Reels are skipped with a reason.
No. It reads only what Instagram shows any logged-out visitor of a public Reel. No login, cookies or API key, so your account is never involved.
From the Reel's audio, transcribed on your machine with Moondream's Parakeet Redux, a 178 MB version of NVIDIA's Parakeet TDT v3. Instagram's own captions are only served to logged-in users. Audio is processed locally and deleted right after.
The 25 European languages of Parakeet v3, with automatic detection: English, German, French, Spanish, Italian, Portuguese, Dutch, Polish, Czech, Slovak, Ukrainian, Russian, Swedish, Danish, Finnish, Estonian, Latvian, Lithuanian, Hungarian, Romanian, Bulgarian, Croatian, Slovenian, Greek and Maltese.
Instagram has no logged-out search, so the skill searches the web for pages that embed or list Reels on your topic, such as articles, roundups and newsletters, and reels_links collects the Reels from them. You can always paste your own links.
--setup creates a private Python environment in ~/.cache/reels-research-mcp with Moondream's Photon runtime (it brings PyTorch, about 1 GB in total) and downloads the 178 MB model. yt-dlp and ffmpeg are needed for Reel data and frames.
No. It is an independent open-source project. Use it in line with Instagram's terms and respect creators' rights when you quote or reuse their content.
Free, open source, no login. Claude Code, Codex and any MCP client.
Researching YouTube too? YouTube Research MCP