Dynamic abstract 3D scene of vivid coral and teal

Why We Built Molten: Private AI That Runs on Your Devices, Your Mac, and Your Rules

Molten 2.0 is now a free download. Here’s why it exists, how it keeps your data yours, and the setups I actually use every day.

The problem: my most useful documents are the ones I’d never upload

The documents that would get the most out of AI are often the ones I’m least willing to hand over. My CVs, cover letters and job-hunting notes. Client briefs and contracts. Financial paperwork. These are exactly where a smart assistant could save me hours, and exactly where I don’t want a copy sitting on someone else’s servers, feeding someone else’s training pipeline, or exposed in someone else’s breach.

So I set out to build the tool I wanted for myself, together with my fellow developer Jack Wang. Molten: Local AI & Doc Chat (Molten for short) is a local-first AI app for Mac, iPhone and iPad. It connects directly to AI you control: Ollama on your own Mac, a fast Apple Silicon runtime like Swama or oMLX, Apple’s on-device model, or any OpenAI-compatible service you choose to trust. Molten is only the client. There’s no Molten server between you and your model.

What Molten is and isn’t

Molten is a chat and knowledge app that talks straight to your AI backend. It isn’t an AI provider, and it doesn’t route, log or analyse your conversations. Our privacy promise is simple: on your devices and your own iCloud — no Molten servers, no account. Our App Store privacy label says Data Not Collected, and that’s literal. The only new network activity in 2.0 is iCloud sync (to your own account) and a one-time download of the embedding model.

You don’t have to take our word for it, because Molten is open source. Every line is on GitHub, so you or anyone you trust can check that it never phones home.

One honest caveat: Molten sends your prompts wherever you point it. If that’s Ollama on your Mac, nothing leaves your network. If it’s a third-party API, that provider sees what you send. You decide, per conversation, and that’s the point.

New in 2.0: your documents join the conversation

The headline feature is Knowledge, which turns your folders into a private, searchable knowledge base. Drop PDFs, text, Markdown, RTF, HTML, DOCX or images into the shared Molten folder in iCloud Drive. Your Mac does the heavy processing once (OCR for scanned PDFs, chunking, embedding), and cited answers follow you to iPhone and iPad. Every answer points back to the document it came from, so you can check it instead of trusting it blindly.

Also free for everyone: search across conversations, export (Markdown, JSON or zip) and import, auto-generated titles, and a context meter that shows how much of the model’s memory you’re using.

If your iCloud data needs end-to-end encryption too, turn on Apple’s Advanced Data Protection. Then even Apple can’t read your synced files.

How I actually use it: the “split brain” setup

The design idea behind Molten is that your devices should share the work sensibly.

At home, the Mac does the heavy lifting. I have a Mac that runs 24/7 with larger open models like Qwen and Gemma, through Ollama or an MLX runtime. It also does all the document processing for Knowledge. That’s where long documents, big context windows and serious reasoning happen.

On the go, the phone uses the Mac remotely. With a secure tunnel (I wrote about doing this with Caddy), my iPhone talks directly to my home Mac from anywhere. A private mesh VPN like Tailscale is another popular option. You join your phone and Mac to the same private network and point the client at the Mac’s private address, so nothing is exposed to the public internet (Source).

For quick tasks, on-device is enough. Molten also supports Apple’s on-device Foundation Model. It has limits: Apple documents a context window of 4,096 tokens per session (Source), so it isn’t the tool for a 40-page contract. But for rewording a message, summarising a short note, or pulling a few facts out of a snippet, it’s instant, free, and works in airplane mode.

Use cases where local AI makes sense

Job hunting. Keep every CV version, job description and interview note in one Knowledge folder. Ask “which of my projects best fits this role?” or “draft a cover letter using my 2023 achievements,” without putting your career history into a cloud service.

Client and commercial work. Consultants, freelancers and agencies often work under NDA. Query briefs, proposals and meeting notes locally, so client information never reaches a third party you haven’t disclosed.

Triage locally, escalate deliberately. This is my favourite workflow. A local model does the first pass: summarise, extract, and find what matters. Then I check whether anything personal or commercially sensitive is in the result. Only if it’s clean, or after I redact it, do I send it to a frontier cloud model for deeper analysis. Local models are also good at the redaction step itself. Researchers have used locally run open models to anonymise sensitive medical reports, precisely so the raw data never goes to the cloud (Source).

Health and family records. Lab results, prescriptions, insurance letters and school reports. Ask questions across years of paperwork without uploading your family’s medical history anywhere.

Personal finance and tax. Statements, receipts and tax returns: “how much did I spend on contractors last year?” These documents are both useful and sensitive.

Contracts and leases. Find notice periods, renewal clauses and obligations, with citations back to the exact page. (Still have a lawyer review anything important.)

Journaling and reflection. Private writing stays private when the model that reads it lives in your house.

Offline and travel. On a plane, or somewhere with poor or untrusted networks, the on-device model and your synced Knowledge base keep working.

Journalists, researchers and anyone with sources to protect. Interview transcripts and source documents are exactly what shouldn’t go to a third-party server.

Practical tips for a good local setup

Start with what you already have. Any Apple Silicon Mac can run useful models, and a Mac mini makes a quiet, low-power home AI server. Set it to stay awake, or to wake for network access, so your phone can reach it. If you’re new to this, start with my guide on how and why to run Ollama on your own Mac.

Match the model to the job. Small models handle rewriting and short Q&A. Mid-size models like Qwen or Gemma on a Mac handle long documents and reasoning. Use cloud models only when you’ve decided the data is safe to share.

Try a faster runtime. On Apple Silicon, MLX-based servers like Swama can be noticeably faster than general-purpose runtimes, and Molten works with either (comparison here).

Lock down remote access. Use HTTPS and an API key, or a private mesh network, and never expose a raw Ollama port to the internet.

Protect your sync. Turn on Advanced Data Protection if you sync sensitive documents through iCloud.

Curious how document Q&A works under the hood? My earlier local RAG chatbot tutorial walks through the same idea that Knowledge now packages up for you.

Built in the open, with AI’s help

Molten is built by a two-person team, Jack and me, working with a set of AI coding agents that help us write, test and review code. We’re open about that, and every line they help produce is public for inspection. The agents work on our code, never on anyone’s data. Molten itself has no servers. It feels right that a privacy-focused AI app is one you can audit yourself, and AI-assisted development is how a small team can ship a Mac, iPhone and iPad app.

Why there’s a subscription

Molten doesn’t sell your data. There isn’t any data to sell. So we fund development the old-fashioned way. The core app is free, including chat with your own models, search, export, auto-titles and the context meter. Molten Plus unlocks the full Knowledge feature (beyond 10 documents, multiple knowledge bases, and iCloud sync) for US$2.99/month, US$19.99/year with a one-week free trial, or a one-time US$49.99 lifetime purchase.

That money pays for a long roadmap: more on-device models for iPhone and Mac, smarter routing between devices, and possibly more platforms later.

If you bought Molten before 2.0, thank you. You’re an Early Supporter, and you get Molten Plus forever: every Plus feature we ship, at no cost and with no account or restore needed.

What’s next

We’re working on smarter device routing: if your home Mac is online with Ollama, Molten should use it automatically, and fall back to a capable on-device model when it isn’t. We’re also adding more small models that fit comfortably on iPhone, so you can choose a capable assistant that never leaves your pocket. Mac and iOS stay the focus for now. Android is the obvious next candidate.

Try it

Your data should work for you without working for anyone else. Download Molten: Local AI & Doc Chat free on the App Store, point it at your own AI, and drop a few documents into the Knowledge folder. Want to see how it works, or contribute? Read the code on GitHub.

Molten is built by Edward Tsang and Jack Wang at OnDemandWorld. If it’s useful to you, a star on GitHub or an App Store review helps a small team more than you’d think.

Privacy Policy · Terms & Conditions