What "Private" Actually Means in the Cloud
Before installing anything, it helps to know precisely what you are avoiding — because the three big assistants have genuinely different policies, and lumping them together as "they read your chats" is lazy.
Google Gemini is the strictest case. Google's own privacy documentation states that a subset of conversations is reviewed by human reviewers, including trained service providers, and that reviewed chats can be retained for up to three years — a retention window that survives deleting the chat in your history. The "Keep Activity" setting is on by default, with an 18-month auto-delete. Switch it off and chats are kept 72 hours and are not used for training.
ChatGPT uses consumer conversations to improve models by default, with an opt-out at Settings → Data Controls → "Improve the model for everyone." Turning it off keeps your history but stops the training use; Temporary Chat does both at once.
Claude is the outlier in the user's favour: Anthropic's privacy centre describes model training on consumer chats as opt-in, not opt-out — it happens only if you actively allow it — and Incognito chats are excluded even then. Conversations flagged for safety review can still be used regardless of the setting.
So the honest version is: the cloud is configurable, and Claude's default is already reasonable. Local AI is for the cases where configuration is not enough — client data under NDA, medical or legal notes, unreleased financials, source code you do not own — where the requirement is that the text never leaves the machine at all, and where you would rather not have to trust a settings toggle to stay where you left it.
The Four Tools Worth Installing
Ollama Best all-rounder
Ollama is MIT-licensed and free, runs on macOS, Windows, Linux and Docker, and is the closest thing this space has to a standard: most other local AI apps can use it as their engine. Its own description of the model is blunt — nothing you run locally ever leaves your machine.
One thing changed in 2026 that older articles miss: Ollama now also sells a Pro tier at $20/month with $60 of credits for cloud-hosted models running in the US, Europe and Singapore. That does not affect the local product — local models remain free and unlimited — but it means "I use Ollama" no longer automatically means "nothing leaves my machine." Check which model you selected before you paste anything sensitive.
LM Studio Free at work
The friendliest graphical option, and the one with the most useful licensing news for anyone in a company. Since 8 July 2025, LM Studio no longer requires a commercial licence for use at work — no form, no sales call, no per-seat cost. The paid Enterprise plan exists only for organisations that want SSO, model and MCP gating, and private collaboration.
It runs MLX and llama.cpp under the hood, downloads models from inside the app, and exposes a local server so other software can talk to it. Same caveat as Ollama: LM Studio also offers optional cloud access to frontier models under a stated zero-data-retention policy. Useful, but it is a different privacy posture from local inference, and you choose which one you are using each time.
Jan Fully offline
Jan is Apache 2.0 licensed and describes itself as an open-source ChatGPT alternative that runs 100% offline. Builds exist for Windows, macOS and Linux (deb, AppImage and Arm64). Of the four, it is the one with the least ambiguity about what it does: an offline chat app, with no paid tier bolted on.
It also publishes clear minimum specs, which is rarer than it should be — see the hardware section below.
AnythingLLM Best for documents
MIT-licensed, free, and available for macOS (Apple Silicon and Intel), Windows (x64 and ARM) and Linux. Its pitch is the one that matters for private work: your models, documents and chat history stay on your machine, nothing phones home, and no account is needed — a one-click install with no signup at all.
Its real differentiator is document chat. If your actual job is "ask questions about these 40 contracts without uploading them anywhere," AnythingLLM is built for that, and it can point at Ollama or LM Studio as the engine rather than replacing them. There is a separate paid cloud product for teams; the desktop app is not it.
Worth naming a fifth: GPT4All from Nomic runs on Windows, macOS and Linux with local document chat and states that no data leaves your machine. It is a reasonable option, though its own site does not clearly publish the app's licence, so verify that before deploying it somewhere it matters.
The Real Cost: RAM, Not Dollars
Every tool above is free. The price is paid in memory, and this is where most "just run AI locally" articles go quiet. Jan publishes the clearest minimums for a decent experience:
- 8 GB RAM — 3-billion-parameter models
- 16 GB RAM — 7B models
- 32 GB RAM — 13B models
Read that as a floor, not a target, and note what it implies. On a typical 16 GB laptop you are running a 7B model — competent at summarising, rewriting, extracting fields from text and answering questions about a document you supply, and visibly weaker than a frontier cloud model at multi-step reasoning, long-context work and anything needing current facts. Being honest about that gap is the whole point: local AI is not a free replacement for ChatGPT or Claude across the board, it is a genuine replacement for the subset of tasks where the input is confidential and the task is mechanical.
Speed is the other trade. A model running on CPU alone will feel slow enough to change how you use it — you will batch questions rather than chat. A discrete GPU or Apple Silicon with unified memory changes that completely. If you are unsure whether a document will even fit in a small model's context window, our Token Counter gives you the size before you start, and AI Text Cleaner strips the invisible characters and formatting artefacts that waste that context when you paste from a PDF.
The Catch Nobody Mentions: Model Licences
The app being open source does not make the model open source. These are separate licences, and for business use the model's terms are the ones that bite.
Llama is the clearest example. Meta's community licence requires a separate licence from Meta if your product had more than 700 million monthly active users in the preceding calendar month — irrelevant for almost everyone, and the reason people wave the issue away. The requirement that does apply to everyone is in the same document: you must prominently display "Built with Llama" on your website, interface, or product documentation if you build on it, and any model you derive and distribute must have a name beginning with "Llama."
Gemma is split. Google moved Gemma 4 to a standard Apache 2.0 licence, while Gemma 1 to 3 and the specialised variants stay under a custom Google licence: commercial use is allowed, but subject to the Gemma Prohibited Use Policy and its distribution conditions. Google claims no rights over the outputs you generate in either case.
The practical rule: if you are chatting privately on your own laptop, none of this touches you. If you are shipping something built on a local model, open the model card and read its licence before the app's — that is the one with the obligations in it.
Comparison Table
| Tool | Licence & Cost | Best For | Watch Out For |
|---|---|---|---|
| Ollama | MIT, free; local models unlimited | The engine other apps plug into | Also sells $20/mo cloud models — check which you picked |
| LM Studio | Free at work since 8 Jul 2025, no licence needed | Teams wanting a GUI without procurement | Optional cloud models are a different privacy posture |
| Jan | Apache 2.0, free | Strictly offline chat, no ambiguity | Smaller ecosystem than Ollama |
| AnythingLLM | MIT, desktop free, no account | Asking questions about private documents | Separate paid cloud product for teams |
| GPT4All | Free; app licence not clearly published | Local document chat, simple setup | Verify licensing before company deployment |
| Cloud (Gemini) | Free tier | Hardest questions, current facts | Human review; reviewed chats kept up to 3 years |
| Cloud (Claude) | Free tier | Best cloud default for privacy | Training is opt-in, but safety-flagged chats are an exception |
Which One Should You Use?
- You want one install and the least fuss: LM Studio. Graphical, downloads models for you, and free at work with no licence to request.
- You want the standard everything else builds on: Ollama. MIT, scriptable, and the engine Jan and AnythingLLM can both point at.
- Your requirement is literally "must not touch the internet": Jan. Apache 2.0, offline by design, nothing else attached.
- Your files are the point: AnythingLLM. No account, nothing phones home, built for chatting with your own documents.
- You have 8 GB of RAM: stick to 3B models and expect summarising and rewriting, not reasoning. If the task is hard and the data is not sensitive, a cloud free tier is the better tool — our comparison of ChatGPT, Claude and Gemini covers which is strongest at what.
- You are automating rather than chatting: a local model behind a self-hosted workflow tool is the private version of an AI agent — see our guide to free AI agent tools for small business, where n8n self-hosted plays the same role.
For everything that does not need a language model at all — counting, formatting, cleaning, converting — a browser tool that runs on your own machine is faster than any chatbot, local or cloud, and equally private.