Open source, self-hosted, or local-only: three things people mean by private

Ask for a private AI notetaker in any technical forum and you will get three kinds of answer: run something open source, self-host it, or use a tool that keeps everything on the machine.

They get offered as if they are the same recommendation. They are three different properties, and you can have any one without the other two. Picking the wrong one is how people end up with a tool that satisfies their principles and not their actual requirement.

The three properties

Open source means you can read the source code. That gives you auditability: you or someone you trust can verify what the software does rather than believing a marketing page. It says nothing about where data goes. An MIT-licensed app configured to call OpenAI ships your transcript to OpenAI exactly like a proprietary one.

Self-hosted means the server is yours. You run it on your own hardware or your own cloud account, so your organisation controls the data at rest and the access to it. There is still a server, and someone in your organisation still administers it.

Local-only means there is no server. The processing happens on the machine in front of you, and the network is not part of the pipeline at all. Turn off wifi and it still works.

You can have open source that is not local. You can have local that is not open source. And self-hosted sits in between: your infrastructure, but still infrastructure.

What the open source options actually do

Meetily is the most visible project in this space, MIT-licensed with an active community. It captures system audio without a bot and runs summarisation either through local models via Ollama or through your own cloud API keys. That last part is the important detail: the privacy outcome depends on how you configure it. Point it at Ollama and nothing leaves your machine. Point it at an API key and your transcripts go to that provider.

Setup requires Node.js, Python and FFmpeg. That is the honest trade for owning the whole stack, and it is a real barrier if you are not comfortable in a terminal.

Anarlog, previously released as Hyprnote, is a local-first open source notetaker backed by Y Combinator. It listens to device audio without sending a bot, transcribes on the machine with Whisper-class models, and keeps app data in a local SQLite database. There is an optional self-hostable server component for teams that need collaboration or admin controls, which is a clean way to separate the individual case from the organisational one.

Whisper itself is not a notetaker, it is the open source speech recognition model that most of these tools use underneath, including ours. If someone tells you to “just use Whisper”, they mean assembling the pipeline yourself: capture, transcription, then a summarisation step you provide.

Which property do you actually need?

You need open source if the requirement is verification. A security team that must audit what runs on managed laptops, a threat model where you cannot take a vendor’s word for anything, or an organisational policy that mandates reviewable code. This is a legitimate and fairly common requirement, and if it is yours, nothing proprietary will satisfy it. Including us.

You need self-hosting if the requirement is organisational control. Data residency in a particular jurisdiction, an existing compliance regime built around infrastructure you administer, or a team that needs shared access to notes without a third-party vendor holding them.

You need local-only if the requirement is that the content never exists anywhere else. Privileged conversations, clinical sessions, pre-announcement deal terms, source protection. Here the question is not who can audit the code or who administers the server. It is whether a copy exists off your machine at all.

Where we sit, stated plainly

Private Notetaker is local-only and not open source. You cannot read our code. That is a real limitation and if auditability is your requirement, Meetily or Anarlog are better answers and we would rather point you at them than pretend otherwise.

What we offer instead is that the thing works out of the box: no Node.js, no Python, no Ollama, no API keys to configure and no way to accidentally configure it into sending your meetings somewhere. Recording, transcription and summarisation all run on your machine, and the airplane-mode test passes because there is no network path to enable.

It is a one-time $149 licence or $15 a month, versus free for the open source options plus your time. For a lot of people the setup time is the expensive part. For others, the ability to read the code is worth every hour. Both are correct.

How to verify any of these claims

The tests are the same regardless of licence:

1. Turn off wifi and run a meeting. If transcription and summarisation both complete, the pipeline is genuinely local. If either queues, something is going out.
2. Check the configuration, not the README. For the configurable tools, the privacy outcome depends entirely on which backend you selected. Open source with an OpenAI key is not private.
3. Watch the network. Little Snitch on macOS or an equivalent shows you what a process actually connects to, which is the only answer that is not somebody’s claim.

We wrote up the full version of these tests, including what the models cost in disk and battery, in local AI transcription on a Mac. For a comparison of the on-device options including several that are not ours, see the best local AI notetakers for Mac.

Project details checked 5 September 2026. Open source projects move quickly, so verify licences and capabilities against the repositories before relying on them.