Reel Digest

CI MIT License

An AI agent that watches the Instagram reels you save and writes a morning digest of workflows you can actually build. Send automation and productivity reels to an Instagram group chat. Every morning, Reel Digest runs Claude Code headless with an Instagram MCP server, downloads each new reel, transcribes it locally with Whisper, reads its keyframes, finds the real links (repos, plugins, MCP servers, product pages), and writes a step-by-step digest tailored to your own goals. It runs unattended on your Mac. Transcription is local; the only service that sees a reel's content is Claude Code, the agent you already use.

• Saved reels you never rewatch? Each one becomes a written summary with the exact prompt, tool names, settings and order of steps.
• "Link in bio" hunting? It finds the repo, plugin or product page from the caption, the screen, the transcript or one web search, and never invents a URL.
• Too many ideas, no plan? One "today's build" per morning, picked against your goals, with steps concrete enough to follow without the reel.
• Deadlines buried in videos? Time-sensitive items (grants, credits, launches) are flagged first.

_{A digest page, rendered from the fictional sample digest.}

Contents

· Features · How it works · Getting started · Day to day · FAQ · Project structure · Status and limitations · Contributing · License

Features

• Reads one DM group. You pick the group once. Each run scans its latest 100 messages for shared reels and posts it hasn't processed yet.
• Watches each reel. It downloads the video, transcribes it locally with faster-whisper, and pulls four keyframes so the agent can read on-screen text, URLs and UI screenshots. A vocabulary hint keeps tool names like "n8n" spelled right. Photo posts are read directly.
• Finds the real links. Repos, plugins, MCP servers and product pages come from the caption, the screen, the transcript, or one web search. It never invents a URL: a missing one is marked "link not found", and link shorteners are marked "unverified".
• Writes a digest tailored to your goals, which you describe in context.md:
  • a TL;DR;
  • flags for time-sensitive items, deadlines first;
  • today's build: one workflow, with steps concrete enough to follow without the reel;
  • for each reel: what it shows, the author's steps, your adapted steps, effort and value, and a verdict (Build, Act, Backlog or Skip);
  • the top 5 from a running backlog (backlog.md). Tick an item [x] and it is never suggested again.
• Delivers it. It renders a standalone HTML page (light and dark mode, click-to-zoom images), shows a macOS notification and opens the page in your browser.
• Runs unattended. It runs at 7:47 every morning, catches up after sleep or a shutdown, waits for the network, runs once a day, and stops after 60 minutes.
• Copes with a laptop's day. It doesn't start during a lid-closed "dark wake" (it waits until someone is at the Mac), keeps the Mac from idle-sleeping mid-run, and if a run fails, retries once when the Mac is in use again. A failed run never blocks the next one.
• Remembers progress. state.json records every processed reel, so each one is handled once. A run takes at most 12 reels; the rest wait for the next morning.
• Cleans up. It keeps only the images the digest embeds and deletes videos and transcripts after a good run. Downloads from a failed run stay for 3 days, so a retry can reuse their transcripts.
• Locked down. The agent can read the group and download reels, but every send, like, delete, mute and mark-seen tool is denied, file edits are limited to three paths, and only one shell command is allowed.

How it works

flowchart LR
    A["launchd<br/>7:47 or login"] --> B["run_digest.sh<br/>waits for you + network"]
    B --> C["claude -p<br/>DIGEST_PROMPT.md"]
    C -->|"list and download reels"| D["instagram_dm_mcp<br/>(your DM group)"]
    C -->|"each video"| E["process_reel.py<br/>keyframes + transcript"]
    C -->|"find links"| F["web search"]
    C --> G["digests/DATE.md<br/>backlog.md, state.json"]
    G --> H["finalize + render HTML"]
    H --> I["notification,<br/>page opens in browser"]

1. run_digest.sh starts, from launchd or by hand. It skips the run if today's digest already exists, checks that the Instagram session and config.json are in place, waits until someone is at the Mac (screen on and keyboard or trackpad used in the last 10 minutes; a run you start by hand passes at once), and waits up to 2 minutes for instagram.com to respond.
2. It runs Claude Code headless (claude -p) with DIGEST_PROMPT.md, filling in today's date and the folder path, under caffeinate so the Mac doesn't idle-sleep. The run is locked down:
   • only the Instagram MCP server from .mcp.json is loaded (--strict-mcp-config);
   • the only Instagram tools allowed are list_messages_compact and download_shared_post_from_message. The send, like, delete, mute and mark-seen tools are denied;
   • edits are limited to digests/, state.json and backlog.md, and the only shell command allowed is tools/process_reel.py;
   • the prompt tells the agent that captions, transcripts and web pages are data written by strangers, never instructions.
3. The agent picks reels. It lists the group's latest messages and takes reels whose shortcode isn't in state.json: new ones oldest first, then older leftovers newest first. Text messages next to a reel are read as notes on why it was saved. For each reel it:
   • downloads the media into media/<date>/;
   • runs tools/process_reel.py, which gets the duration with ffprobe, extracts 4 evenly spaced keyframes (480 px wide) with ffmpeg, and transcribes the audio with the faster-whisper small model on the CPU (int8, with voice activity detection). Results are cached next to the video;
   • reads 2 to 4 keyframes and looks up the links.
4. It writes the digest to digests/<date>.md, updates backlog.md and state.json, and ends with one line: OK: <n> new reels · <f> flags · today's build: <name> or ERROR: <reason>.
5. run_digest.sh finishes the job. On OK, finalize_digest.sh copies the embedded images into digests/img/<date>/ and deletes that day's downloads. tools/render_digest.py renders the HTML page; raw HTML in the Markdown is escaped, because captions are written by strangers. digests/latest.html is pointed at the new page and a notification appears. On ERROR or a timeout, it waits until the Mac is in use again and retries once (not for an expired session); if that fails too, the notification gives the reason and the failed digest is moved to logs/ so the next login tries again. If the Instagram session has expired, the digest says so and tells you to run ./login_once.py.

Tech stack

• Claude Code CLI, run headless (claude -p) as the agent, with tool allow and deny lists
• MCP: instagram_dm_mcp (my fork of trypeggy's server), built on instagrapi
• Python 3.11 or 3.12 with uv: faster-whisper for local speech-to-text, markdown-it-py for the HTML page
• ffmpeg and ffprobe for keyframes and video duration
• zsh, launchd, caffeinate and osascript on macOS for scheduling, keeping awake and notifications

Getting started

Prerequisites

• macOS on Apple silicon or Intel. The scripts use launchd, osascript, caffeinate, pmset, open and BSD sed. It runs every morning on an M5 Pro MacBook Pro.
• Claude Code, installed and logged in, with claude on your PATH. Each run uses your Claude Code plan or API key.
• uv, plus ffmpeg and ffprobe on your PATH (brew install uv ffmpeg).
• An Instagram account that is a member of the group you send reels to. A secondary account is safer; see Status and limitations.

1. Clone this repo and the Instagram MCP server

The Instagram MCP server is a separate project. Clone my fork of it into instagram_dm_mcp/ inside this folder. That path is gitignored, and the scripts expect the server there.

git clone https://github.com/hanshika0292/reel-digest.git
cd reel-digest
git clone https://github.com/hanshika0292/instagram_dm_mcp.git instagram_dm_mcp
uv sync --project instagram_dm_mcp

Use the fork, not upstream. Upstream's pinned instagrapi is rejected by Instagram, it crashes on shared reels, and it lacks the list_messages_compact tool the digest uses. The fork fixes these; the fixes are also open as upstream PRs #14–#17.

2. Add credentials and install the Python tools

cp .env.example instagram_dm_mcp/.env && chmod 600 instagram_dm_mcp/.env   # then fill it in
uv sync --project tools                                                    # faster-whisper, markdown-it-py

To run the tests (they mock ffmpeg and Whisper, so they need neither): uv run --project tools pytest tools/tests

The project reads these environment variables from instagram_dm_mcp/.env:

Variable	What it is for
INSTAGRAM_USERNAME	The Instagram account in the group. It also names the session file.
INSTAGRAM_PASSWORD	The MCP server refuses to start without it. login_once.py --password logs in with it.

Claude Code uses its own login, so the project reads no API keys. The first transcription downloads the Whisper small model.

3. Log in and pick the group

./login_once.py

Run it from the project folder. Instagram currently rejects password logins from scripts (instagrapi issues #2807 and #2881), so by default the script walks you through copying the sessionid cookie from an Incognito browser login. It saves the session to instagram_dm_mcp/<username>_session.json (chmod 600), lists your 40 most recent DM threads with groups first, and writes the chosen group's ID and title to config.json. ./login_once.py --password tries a password login instead.

4. Describe your goals

cp context.example.md context.md

Edit context.md. Give each goal a short bold label, and mark time-sensitive goals "(always flag)". The digest tailors every plan to this file.

5. Point the MCP config at this folder and run once

sed -i '' "s|/path/to/reel-digest|$PWD|" .mcp.json
./run_digest.sh

This writes your local path into .mcp.json, which is tracked, so don't commit that change. The first run handles up to 12 reels from the group's latest 100 messages.

6. Run it every morning at 7:47

com.reeldigest.morning.plist is a launchd template. It runs the digest at 7:47 every day and at every login. If the Mac is asleep at 7:47, launchd runs it on wake, and the script waits until you're actually using the Mac. The login run covers a Mac that was shut down. Extra runs do nothing, because run_digest.sh skips a day that already has a digest.

mkdir -p logs
sed "s|/path/to/reel-digest|$PWD|g" com.reeldigest.morning.plist > ~/Library/LaunchAgents/com.reeldigest.morning.plist
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.reeldigest.morning.plist

To change the time, edit Hour and Minute in the installed plist, then run bootout and bootstrap again. To stop it, run launchctl bootout gui/$(id -u)/com.reeldigest.morning.

Day to day

To	Do
Retry after a failed run	./run_digest.sh --force
Fix "Instagram session needs a refresh"	./login_once.py
Watch a different group	./login_once.py, then answer y to "Change it?"
Get the notification without the browser tab	Add "open_report": false to config.json
See what happened in a run	logs/<date>.log (and logs/<date>.out for the agent's output)
Ask questions about the group	Open Claude Code in this folder. It loads the Instagram server from .mcp.json, with all of its tools, including send.

FAQ

How do I turn saved Instagram reels into notes automatically? Send them to an Instagram group chat you're in, point Reel Digest at that group once with ./login_once.py, and it writes a digest of the new ones every morning: summary, links, steps and a verdict.

Does it post, like or send anything on my Instagram account? No. The agent is only allowed to list the group's messages and download shared posts. The send, like, delete, mute and mark-seen tools are explicitly denied on the command line.

Is this allowed by Instagram? It uses an unofficial API (instagrapi), which can conflict with Instagram's Terms of Use and can get an account challenged or rate-limited. It reads only a group you're a member of, once a day. Use a secondary account if that matters to you. See Status and limitations.

Do I need an OpenAI or Anthropic API key? No separate key. The agent is your installed Claude Code, using its own login (a Claude plan or an API key you've already set up). Transcription runs locally with Whisper, so no audio is sent to a speech-to-text service.

What does it cost to run? The Claude Code usage of one run a day (up to 12 reels and a few web searches each). Whisper, ffmpeg and everything else run locally for free.

Does it work on Apple Silicon Macs (M1–M5)? Yes. It runs daily on an M5 Pro, and the Python tools' lockfile covers both Apple Silicon and Intel Macs.

What if my Mac is asleep or the lid is closed at 7:47? launchd runs it when the Mac wakes, but the script waits until someone is actually using the Mac, so it doesn't crawl through a lid-closed "dark wake". If the lid closes mid-run, it retries once when you're back.

Can a malicious reel hijack the agent (prompt injection)? The risk is reduced, not removed. Captions, transcripts and web pages are treated as data, the agent can't send anything or run other commands, and it can only write to three paths. It can still search and fetch web pages.

Project structure

reel-digest/
├── run_digest.sh                 entry point: checks setup, runs Claude Code headless, retries once, finalizes, notifies
├── wait_for_user.zsh             waits until someone is at the Mac (no runs during a lid-closed dark wake)
├── DIGEST_PROMPT.md              the agent's instructions: which reels to take, how to find links, digest format
├── finalize_digest.sh            keeps the images a digest embeds, deletes that day's downloads
├── login_once.py                 Instagram session login and group picker (writes config.json)
├── tools/
│   ├── process_reel.py           video → keyframes (ffmpeg) and transcript (faster-whisper)
│   ├── render_digest.py          digest Markdown → standalone HTML page
│   ├── tests/                    pytest tests for both tools (ffmpeg and Whisper are mocked)
│   ├── pyproject.toml
│   └── uv.lock
├── context.example.md            template for context.md (your goals)
├── examples/sample-digest.md     a short, fictional digest
├── docs/                         screenshots
├── com.reeldigest.morning.plist  launchd template: 7:47 daily and at login
├── .mcp.json                     Claude Code MCP config for the Instagram server (template path)
├── .env.example                  Instagram credentials, copied to instagram_dm_mcp/.env
├── THIRD_PARTY_NOTICES.md        what it builds on, and their licenses
└── LICENSE                       MIT

These are created locally and gitignored: instagram_dm_mcp/ (the cloned server, its .env and the session file), config.json, context.md, state.json, backlog.md, media/, digests/ and logs/.

Status and limitations

• This is a personal tool that I run every morning on my own Mac. It only runs on macOS and isn't packaged as an app. Tests cover tools/ only; the shell scripts and the agent run are tested by hand.
• It uses an unofficial Instagram API. The MCP server is built on instagrapi, which imitates the Instagram mobile app to call Instagram's private API. It can break whenever Instagram changes something, and automated access like this can conflict with Instagram's Terms of Use. Instagram may challenge, rate-limit or lock the account. It is meant for your own account only, reading a group you are a member of. Keep it to one run a day.
• Sessions expire. When one does, you copy a fresh sessionid by hand; the unattended run can't log in.
• It depends on my fork of a third-party MCP server (see step 1). The original upstream server won't work with the digest.
• Reels are written by strangers, so prompt injection is a real risk. The prompt says to treat them as data, and the tool lists block sending, limit writes to three paths and allow one shell command. The agent can still search and fetch web pages, so this reduces the risk but doesn't remove it.
• It only sees the group's latest 100 messages, and only 4 keyframes per video, so on-screen text shown between keyframes can be missed.
• Runs are slow and use Claude Code usage: up to 12 reels, local transcription on the CPU, and a 60-minute limit.
• Instagram is a trademark of Meta Platforms, Inc., and Claude Code is a product of Anthropic. This project isn't affiliated with or endorsed by either.

Contributing

Issues and pull requests are welcome, especially fixes for Instagram API changes and improvements to DIGEST_PROMPT.md. Run uv run --project tools pytest tools/tests before opening a PR. Never include a real .env, session file, context.md or digest in a PR. By contributing, you agree that your contribution is licensed under the MIT License.

Credits

Built on instagram_dm_mcp by Gala Labs (via my fork), instagrapi, faster-whisper and OpenAI's Whisper model, markdown-it-py, FFmpeg and Claude Code. See THIRD_PARTY_NOTICES.md.

License

MIT. Copyright (c) 2026 Hanshika Gupta.