Private meetings with local AI: setup and what stays on your computer
You can run Natively so that meeting content stays on your computer: an on-device speech model transcribes the call, a model you run in Ollama writes answers and notes, and a local embedding model powers search across past meetings. Transcripts, notes and the search index are saved locally and these features keep working without internet, but the app still sends usage statistics (on by default, can be turned off), license checks and update checks whenever it is online.
By Evin JohnPublished 6 min read
Why some meetings should not touch a cloud AI service
Most AI note takers send meeting audio to a speech service and the transcript to a language model, both run by other companies. That is fine for many calls. It is a problem when the conversation is a client's legal strategy, an unreleased financial result, a patient case, or anything covered by a confidentiality agreement that names who may process the data.
Natively already avoids one part of the problem: no bot joins the meeting, because it captures your computer's system audio and microphone directly. The local setup removes the second part. Audio, transcript text and prompts are processed on your machine instead of being sent to a speech or AI provider. For the broader trade-offs, see local AI vs cloud AI assistants.
The three local pieces
- Speech to text: an on-device model. Natively can download Whisper (several sizes, English and multilingual), Distil-Whisper, Parakeet (v2 English, v3 multilingual), Moonshine (English) and Nemotron streaming models. On macOS 26 and later, Apple Speech is also available. With any of these, no audio leaves the computer.
- Answers and notes: a model in Ollama. Natively can use any model you run in Ollama on your own computer, with no API key. Prompts with transcript context go to that local model instead of a cloud provider.
- Search across meetings: local embeddings. The search index is stored locally either way. A local embedding model is bundled as the no-key option; if you choose a cloud embedding provider instead, the text being embedded is sent to that provider.
Set it up step by step
- Install NativelyIt runs on macOS 12 or later and Windows 10 or 11 (x64). Do this while online, because models are downloaded in the next steps.
- Configure AI before your first meetingIf no AI is set up when you start a meeting, a 30-minute free trial begins by itself and uses the hosted Natively API for transcription and answers. Finish the next three steps first if you want your first meeting to be local too.
- Choose a local speech modelIn Settings, choose an on-device speech model and let it download. Sizes range from about 26 MB to about 2.85 GB. Pick a multilingual model if your meetings are not in English.
- Install Ollama and pull a modelInstall Ollama yourself, download a model with it, then select Ollama as the AI provider in Natively. It connects to Ollama's default local address, http://127.0.0.1:11434.
- Keep embeddings localLeave search on the bundled local embedding model rather than a cloud embedding or reranking provider.
- Turn off usage statistics if you wantProduct-usage events are on by default. Switch them off in Settings > General > Advanced.
- Test with the network offDisconnect from the internet, run a short test call, and confirm that the transcript, an answer and the notes all appear.
Hardware and quality, honestly
Local models use your computer's memory, processor and graphics chip. Larger speech models transcribe more accurately but need more memory and keep up with the conversation less easily on older machines; Whisper medium is offered on Apple Silicon only. In Ollama, small models answer more slowly and less accurately than large cloud models, and notes from a small model will miss more nuance. A recent machine with plenty of memory makes the biggest difference.
Start with a mid-sized speech model and a mid-sized Ollama model, run a few internal calls, and move up or down from there. Screenshot questions need a model that can read images, so check what your local model supports before relying on it for slides or documents.
What stays on the computer and what goes online
| Data | Where it goes | When |
|---|---|---|
| Call audio (microphone and system audio) | Stays on the computer; processed by the on-device speech model | Always, in this setup |
| Transcripts, notes, typed notes, screenshots, meeting history | Stay on the computer, in the app's local database | Always |
| Prompts with transcript context | Stay on the computer; sent to your Ollama model | Always, in this setup |
| Search index and embeddings | Stay on the computer with the local embedding model | Always, in this setup |
| Local MCP server and local API | Your machine only (127.0.0.1), read-only | Only if you turn them on |
| Usage statistics: events such as the app opening, a meeting starting, which features ran | Natively's server, linked to a device id. No transcripts, screenshots or typed text | When online, unless switched off in Settings > General > Advanced |
| Analytics from the app window (Google Analytics): app open and close, mode, model used, session length | Google Analytics. No meeting content | When online; separate from the usage-statistics switch in the current version |
| Install ping: install id, version, platform | Natively | Once, after installing |
| License check | Natively's license server and the payment provider | When online, if you use Pro |
| Update check | GitHub | When online |
So the accurate claim is this: in a fully local setup, meeting content stays on the computer, while a small amount of product and license information goes out when a connection exists. Anything you send yourself (a connector, a share link) also leaves the machine, by design. The AI note taker privacy guide covers the questions to ask of any tool.
What stops working without internet
- Cloud speech and AI providers, and the hosted Natively API.
- The free trial, which runs on the Natively API.
- Calendar sync with Google, Outlook or an ICS link.
- Connectors: Slack, Notion, Gmail, HubSpot, Salesforce, Attio, Pipedrive, Affinity, Clearbit, Zapier and webhooks.
- Share links, which publish notes to a page on Natively's server.
- Downloading new speech or Ollama models.
If you have Pro, its templates and modes keep working while you are offline. Pro features that depend on a web search, such as company research, still need a connection.
Fictional example
- Setup
- Whisper Large v3 Turbo for speech, a mid-sized model in Ollama, local embeddings, usage statistics switched off. Wi-Fi off for the meeting.
- During the call
- The client walks through a draft of next year's cash forecast. The accountant asks Natively to list every assumption the client has mentioned so far, and it answers from the live transcript.
- After
- The local model writes notes with action items. The accountant then asks across past meetings which of those assumptions differ from the spring review, exports the notes as PDF and files them in the firm's document system. Nothing goes to Slack or a CRM.
Who this setup is for
- Lawyers, accountants, consultants and advisers bound by client confidentiality.
- Teams in regulated settings whose policy restricts which processors may handle meeting content.
- Engineers discussing unreleased products or internal systems.
- People without API keys who do not want a subscription for transcription and answers.
- Anyone who often works without a reliable connection.
Plans for a local setup
| Scenario | Feature | Plan |
|---|---|---|
| Transcribe on the computer | On-device speech model | Free |
| Answers and notes without an API key | Ollama local models | Free |
| Search and questions across past meetings | Local search index with local embeddings | Free |
| Notes in the General mode | Summary, sections, action items | Free |
| Note templates such as 1 to 1 or Board Meeting | Note templates | Pro |
| Modes other than General | Built-in and custom modes | Pro |
There is more on the product side on the local AI assistant page.
What Natively does not do
- It does not install Ollama or pick a model for you.
- It does not make a small local model as capable as a large cloud model.
- It does not split several voices on the other side of the call by itself. That comes from Deepgram, Soniox or the Natively API, which are cloud services; locally it separates you from the other side.
- It does not run on Linux.
- It does not make every network request stop. Usage statistics can be switched off; license and update checks run whenever the computer is online.
Questions people ask
Can Natively work completely offline?
Transcription, answers, notes and search keep working without internet when you use an on-device speech model, an Ollama model and local embeddings. Calendar sync, connectors, share links, the trial and the Natively API need a connection, and models must be downloaded first.
Does anything leave my computer in a local setup?
Meeting content does not: audio, transcripts, prompts and the search index stay on the computer. When online, the app still sends usage statistics (on by default, can be turned off in Settings), analytics events from the app window, a one-time install ping, license checks and update checks, none of which carry transcripts, screenshots or typed text.
Are local models as good as cloud models for meeting notes?
Usually not. Small local models answer more slowly and less accurately than large cloud models, and results depend on your computer's memory and graphics chip. Test a few internal calls and size the models to your machine.
Where to go next
Natively runs on macOS and Windows. The download is free; Pro and the Natively API are optional.
Facts on this page were last checked on . Product details come from the Natively app itself; if something here is out of date, tell us at hello@heynatively.com.
