Reel Library (instamcp)

You saved hundreds of reels "for later". Later never came. Reel Library watches them for you: it transcribes each one, writes a summary, pulls out every tool, place, product and link, and makes the whole collection searchable from a clean web app or from any AI assistant.

Unofficial, personal-use tool. Not affiliated with, endorsed by or sponsored by Meta or Instagram. It works on reels you saved, using your own login. Automated downloading may conflict with Instagram's Terms of Use: you are responsible for how you use it. Respect creators' rights and don't redistribute downloaded content.

Turn your Instagram saved reels into a searchable, summarized, organized library, with a web app to browse it and an MCP server so any AI assistant can use it in chat. The analysis runs on the AI engine of your choice: Claude, OpenAI, Google Gemini, OpenRouter, Groq, a free local model through Ollama or LM Studio, or any OpenAI-compatible endpoint, with automatic fallback between them.

For every saved reel it:

1. Downloads the video using your Instagram login.

2. Watches it: samples keyframes and transcribes the speech locally (faster-whisper).

3. Understands it with your chosen AI engine (vision + transcript + caption + the names of your saved collections): title, summary, key points, category, content type, topics, tags, on-screen text, and why you probably saved it.

4. Extracts links & resources: every tool, app, product, book, place, recipe, repo or site mentioned in the caption, spoken aloud or shown on screen. URLs are checked, guessed URLs that don't resolve are dropped, and every resource gets a search link.

5. Organizes everything by your Instagram collections and by AI-assigned categories, with full-text search over all of it.

6. Reads every reel with Laya, a fast local decision model, to add signals (actionable? worth revisiting? why you saved it) that drive the Analytics page. You can also ask any question across your whole library, e.g. "Is this an activity to do with a toddler?", and get the reels ranked by the answer in seconds.

See it in action

Search everything you saved, then open a reel for its summary, key points and links

Every tool, place and product mentioned across your reels, in one list

Ask one question across your whole library and get the reels ranked by the answer

Related MCP server: FeedRecall MCP Server

Screenshots

Screenshots use a made-up demo library.

                        ┌──────────────── Docker container (port 8090) ────────────────┐
 Instagram export.zip ─►│ Import ─► SQLite + full-text index ◄─ Analysis worker         │
                        │                 ▲        │             yt-dlp → ffmpeg frames │
 Browser  ─────────────►│  React UI ──► REST API    │             whisper → AI engine   │
 AI assistant (MCP) ───►│  /mcp (MCP over HTTP) ────┘                                   │
                        └──────────────────────────────────────────────────────────────┘
                                         volume: /data (db, thumbnails, settings, models)

Run it with Docker (recommended)

On Windows, see windows/README-WINDOWS.md for a step-by-step guide and one-click start script.

Requires Docker Desktop. Download the source (or the latest release), then:

cd reel-library
docker compose up -d

This pulls the ready-made image (ghcr.io/ic3w4r/reel-library, for Intel/AMD and Apple Silicon), a 1-2 GB download the first time. Open http://localhost:8090. To use another port: INSTAMCP_PORT=9000 docker compose up -d. To build from source instead: docker compose up -d --build.

Then, in the app:

1. Settings → AI engines: pick a provider, paste its API key (local models need none), choose a model (Fetch models lists what your key can use), and click Test. Optionally add a second engine as the fallback.

2. Settings → Instagram login: paste your sessionid cookie (the page shows where to find it) or upload a cookies.txt. Instagram only serves reel videos to logged-in users.

3. Add & analyze: drop in your Instagram data export zip (how to get it is below) and leave Start analyzing right after import ticked. Progress shows in the sidebar.

4. All saved: browse by collection or category, search everything, and click a reel for its summary, key points, links, and transcript. Links & resources lists every link across all reels and can copy them as Markdown.

Everything is stored in the instamcp-data Docker volume, so it survives restarts and rebuilds.

The port is bound to 127.0.0.1 only. The app has no login of its own and stores your API keys, so don't expose it to a network without putting authentication in front of it.

Getting your Instagram export

Instagram → Settings → Accounts Center → Your information and permissions → Download your information → choose your account → Some of your information → tick Saved → download to device, format JSON, date range All time. Instagram emails you when it's ready. Re-importing a newer export later only adds new reels.

You can also add single reels by URL on the Add & analyze page.

AI engines

Settings → AI engines holds any number of engines. The primary analyzes every reel; if it fails because of the account or setup (bad key, no credits or quota, unknown model, server unreachable), the fallback takes over for that reel automatically. If every engine is unusable, the queue pauses with the reason instead of marking your reels failed.

• Pre-filled model names are suggestions. Use Fetch models to pick from what your key can access, and prefer a model that accepts images, since it then sees the video frames.

• The OpenAI-compatible adapter adapts to each endpoint: if strict JSON schemas aren't supported it falls back to JSON mode, then to prompt-only JSON. If the model can't take images it sends the transcript and caption only. Output from any model is normalized to the same categories, types and link kinds.

• Local servers from Docker: presets point at host.docker.internal, which reaches Ollama or LM Studio running on your Mac.

• API keys set as environment variables (see docker-compose.yml) are picked up by the matching engine when you don't type one in the UI.

Analytics and Laya decisions

Laya is a non-generative decision model: given text and a typed question it answers in one forward pass, with no text generation and so nothing to parse or hallucinate. It answers as choice (pick a label), score (a level on a scale) or yes/no (a probability). It runs locally, in 100+ languages.

• Signals. Each reel with text (an AI summary, a caption or speech) gets category, content type, purpose (learn / entertain / inspire / shop), actionable and revisit value. These are computed automatically after each analysis (toggle in Settings), or in bulk with Read N reels on the Analytics page.

• Analytics page. Saves per month, categories, why you save, kinds of reels, how much is worth revisiting, and a per-collection table.

• Ask your library. Ask a yes/no, pick-one or scale question across up to 2,000 reels, then filter by answer and open any reel. Also available to AI assistants as the ask_reels MCP tool.

• Which checkpoint answers. Choice questions run on laya-multilingual, which classified a mixed test set best. Yes/no and scale questions go through Laya's router: English text to the English checkpoint, which separates yes from no more cleanly, and other languages to the multilingual one.

• Limits. Laya's base checkpoints are quick estimators, not experts; its authors note they are weak on complex decisions without fine-tuning. Use the answers to find and sort, and check before relying on one. A reel needs text first, so un-analyzed reels from the export are skipped until they're analyzed.

• Size. Laya brings PyTorch (CPU build in Docker). Its checkpoints (~1–2 GB) download into the data volume the first time it runs; the first run takes about a minute, then it's under a second per reel on a Mac and a few seconds on CPU.

Connect an AI assistant (MCP)

The container serves MCP over HTTP at http://localhost:8090/mcp.

Claude Code

claude mcp add --transport http --scope user instagram http://localhost:8090/mcp

Other MCP clients (Cursor, VS Code, …) take a config like {"mcpServers": {"instagram": {"url": "http://localhost:8090/mcp"}}}. Settings shows both, ready to copy.

Then ask things like:

• "Give me an overview of my saved reels by collection."

• "Find that reel about the garlic pasta and give me the steps."

• "List every AI tool I've saved, with links."

• "Make a 4-week workout plan from my Fitness saves."

MCP surface

Resources: instagram://overview, instagram://reel/{id}, instagram://collection/{name}. Prompts: digest_collection(collection), plan_from_saved(goal).

Running without Docker (development)

Requires uv, ffmpeg and Node 20+ (brew install uv ffmpeg node).

uv sync --extra decisions      # drop --extra decisions to skip Laya / PyTorch
(cd web && npm install && npm run build)
uv run instamcp web            # UI + API + MCP on http://127.0.0.1:8080

For UI work, run uv run instamcp web plus cd web && npm run dev (Vite on :5173, proxied to :8080).

Other commands: instamcp serve (MCP over stdio for Claude Desktop/Code), instamcp import <zip>, instamcp add <url>, instamcp analyze --limit 50, instamcp stats. Outside Docker, the default data folder is ~/.instamcp, and INSTAMCP_COOKIES_FROM_BROWSER=chrome can borrow your Chrome login instead of a pasted cookie.

Configuration

Most settings can be changed on the Settings page; these environment variables set the defaults.

Notes

• Each reel takes about 20–60 seconds. Cost depends on the engine: very roughly 10¢ per reel on Claude Opus 5.5 at medium effort, much less on smaller or "flash"-class models, and nothing on a local model. The first analysis also downloads the speech model (~500 MB, kept in the data volume).

• Claude requests on current models also opt into Anthropic's server-side refusal fallbacks, so a rare model decline is retried instead of failing the reel.

• Photo/carousel posts with no downloadable video are analyzed from the caption when one is known, otherwise marked failed with the reason (retry from Add & analyze).

• Everything stays on your machine except the call to your AI engine (frames + transcript + caption) and link checks. With a local engine (Ollama / LM Studio), nothing leaves your machine except link checks.

Contributing

Ideas, bug reports and pull requests are welcome. See CONTRIBUTING.md for how to run it locally and send changes, and SECURITY.md to report a vulnerability privately.

License

Apache License 2.0. See NOTICE for third-party components. Instagram is a trademark of Meta Platforms, Inc.