Step 1: fetch and listen

The moment a Reel or video lands, fetchers (yt-dlp, OG metadata, readme and custom fetchers, Apify where needed) pull the source. Speech goes through Groq-hosted Whisper transcription; captions, on-screen text, and OCR'd frames are captured separately. Keeping speech, text, and visual context distinct matters — a recipe that says “salt to taste” and a caption that says “add 2 tsp salt” are different kinds of evidence.

Step 2: understand, don't summarize

A wall of text is not understanding. The extraction step returns structure: a concise summary, grounded key points (each traceable to the source), a category, difficulty, estimated minutes, estimated cost, priority, and a coarse project mode — action, understand, or reference. That last field is the honest one: a documentary and a shelf-repair video both get saved, but only one of them deserves a checklist. Non-actionable saves keep their key ideas without a manufactured task list.

Step 3: plan only when a plan exists

For actionable content, task generation produces an ordered list where each step carries a time estimate and, where relevant, a cost — parts, tools, materials. Steps are bounded and re-tried with fallbacks rather than padded with filler tasks. The test for a good generated plan is simple: could you start the first task tonight without rewatching the source?

Step 4: remember, and hand off

Everything lands in a searchable library — projects, summaries, embeddings (pgvector, 1536-dim), a knowledge graph of concepts and tools, and curated memory facts with confidence and decay. From there, two exits matter: search and chat across your own library in the app, and a personal, read-only MCP server so your own agents can read the same context. You can even push a project to a durable handoff queue and let your agent pick it up next session.

The contract

Thirty seconds after the share, you should be looking at a decision: start the first task, study the roadmap, file it as reference, or delete it without guilt. That is the whole product promise. oranoai.com has the details; the app is free on iOS and Android.

What the server does

ORANO sits on your phone's share sheet. Share a Reel, a YouTube video, an article, a GitHub repo, or a PDF, and it builds a structured project: a summary, grounded key points, category, estimated time and cost, priority, and — when the content is actionable — an ordered task list. Learning-type saves can generate a study roadmap; research saves pull sourced context. All of that lives in Postgres with pgvector embeddings.

The MCP endpoint exposes a handful of tools to your own agent — ChatGPT, Claude, Cursor, anything that speaks MCP:

• list your projects, filtered by status

• get a project's summary, structured understanding, resources, research, roadmap, and ordered tasks

• retrieve projects explicitly handed off from the app

• search the library by title, summary, source, or URL

• read curated memory facts with confidence and freshness signals

Why read-only

Three reasons, in order of importance.

1. The failure modes are asymmetric. A read-only integration can fail to help you. A write integration can destroy state you spent months curating — overwriting tasks, duplicating projects, mangling memory facts based on a hallucinated plan. For a personal memory layer, the downside risk of writes dwarfs the upside.

2. Trust is easier to grant than to audit. “Your agent can look at your library” is a decision you make once. “Your agent can change your library” is a decision you re-litigate every time you review a diff. Read-only keeps the mental model simple: the app writes, the agent reads, you decide.

3. The handoff queue is the write path that matters. The flow people actually want is: in a chat on the couch, say “send this project to my agent” — and that is exactly what the durable handoff queue does. The app pushes a project to the queue with your explicit action; the agent picks it up on the next session. The direction of control stays human-initiated, which is where it should stay until agent reliability earns more.

Auth reality check

Authentication is a personal API key with bearer auth scoped to orano:read, not a Supabase session — agents are machines, they should get machine credentials. Rate budget: 240 calls per user per hour, up to 10 active keys. One honest caveat: OAuth discovery metadata exists but the full delegated OAuth flow is not finished, so today a personal key is provisioned manually. When OAuth lands, the read-only default will not change.

Try it

ORANO is free to download on iOS and Android — oranoai.com has the details and the MCP documentation, and the repo orano-mcp-examples has curl examples for every tool. If you build agents, a read-only window into what your user actually saves is the most honest context signal you can get — it is what they chose to keep, not what they remembered to mention.