Bring Your Own Key, Forever

There is a quiet pattern in most AI-flavored software. You sign up. You pay. The tool calls some model on your behalf. The relationship between you and the model provider is mediated, abstracted, hidden. You do not see the prompt. You do not see the token count. You do not know which model version answered. You did not pick that model, and you cannot swap it.

That is the default, and I think it is the wrong default for a tool that reads your meetings.

What BYOK actually means here

Bring-Your-Own-Key, in lognote, means you choose the provider that writes your meeting summaries, and you pay them directly. The tool acts as a thin broker. Your API key sits on your Mac. The request goes from your laptop to the provider you picked, with no middleware in between.

What ships today: local-mlx as the default on-device path (no key required, no network), with bring-your-own-key for direct OpenAI, Anthropic, Azure, or Ollama if you want a different model. If your key is missing or the call fails, the transcript still lands in your note. Only the summary step is best-effort. That is deliberate. The summary is gravy; the transcript is the meal.

Where you set the key

Most users will set this in lognote’s Obsidian settings tab. Open Settings, find the lognote section, and paste your provider details into the fields. The plugin writes them where the recorder reads them, and the next meeting picks them up.

Alpha and power users can also set the same values via shell environment variables or a local config file in ~/.config/lognote/. The Obsidian settings just write to the same place. Either surface is fine; the settings tab is the canonical one, and the env-var route exists because the engine has to be runnable from the command line for the architecture to make sense.

Why local-first summarization is the default

The entire promise of the tool is that your audio stays on your Mac. Transcription already runs there. As of this release, the summary step does too, by default.

If you switch to a cloud provider, the transcript leaves your machine for the duration of the summary call. That is an honest trade you opt into by changing providers. The promise has always been about the audio, not about the summary text; the audio still never leaves. With local-mlx as the default, the trade is now opt-in rather than the only way to get a summary at all.

Local-mlx will not be as polished as a frontier API. The models that fit on a laptop are smaller and less articulate. For 80% of meetings, that is fine. For high-stakes ones, you will reach for a frontier provider and bring your own key to it. That is the right shape of the decision.

The trade-offs, written down

Every provider is a trade. Here is the honest shape.

Local-mlx (the default)

What you get: zero network, zero per-call cost beyond electricity, zero leakage of your transcript to a third party. Picks model size by transcript length (3B for short meetings, 8B / 14B for longer ones), all on-device.
What you give up: quality. The summary will be functional, not brilliant. Action items get captured; subtle nuance sometimes does not.
When to pick it: by default. Always, for personal or low-stakes meetings.

Ollama (self-hosted)

What you get: still no leakage off your machine, but with the freedom to run any model Ollama supports. Useful if you have a beefier desktop on the LAN doing the heavy lifting.
What you give up: the simplicity of local-mlx. You manage the Ollama daemon and the model files.
When to pick it: if you already run Ollama for other tooling, or you want a specific model lognote does not pre-package.

Direct OpenAI / Anthropic

What you get: frontier quality without the Azure layer. Faster setup, less plumbing.
What you give up: the transcript leaves your machine for the duration of the API call. Per-call cost, billed directly to your provider account.
When to pick it: for meetings where the summary will be reused. Client calls that turn into proposals, design reviews that become docs.

Azure OpenAI

What you get: frontier model quality, billed through whatever Azure tenancy you already have, with whatever enterprise governance applies.
What you give up: same transcript-leaves-machine trade as direct OpenAI, plus more plumbing to set up.
When to pick it: if your employer mandates AI calls go through a managed enterprise channel, or you already pay Azure and want one less invoice.

Why this is a permanent choice, not a phase

A statement about the future I’d like to be held to:

lognote will always support BYOK. There will not be a “Pro plan” that quietly hides your API key behind a hosted service. There will not be a future where the only way to get good summaries is to subscribe to lognote’s own AI tier. The provider relationship is yours. That is not a launch promise; it is the architecture.

The reason is simple. The moment a tool inserts itself between you and your model, three things happen, and they all happen quickly. The prompt becomes a trade secret. The model choice becomes a business decision, not a user one. And the tool’s incentives drift toward the cheapest model it can get away with, not the best one for your meeting.

I have used tools that started in the right place and ended in the wrong one. It is not malice. It is gravity. Once a company sits between you and the model, the gravity pulls in one direction. The way to avoid that gravity is to never sit there in the first place.

What this means for product decisions

A consequence of BYOK worth being explicit about: the cost of using lognote, even with a frontier model, is mostly the API bill. If you summarize one meeting a day with a sharp model, you’ll spend a few dollars a month with the provider. If you summarize zero, you’ll spend nothing. The variable cost of the AI sits with the person who actually benefits from it.

This also means lognote can be aggressive about features that route everything through your chosen provider. Re-summarizing with a different prompt, regenerating with a different model, running a one-off “extract decisions” pass on an old transcript: all of that is your provider, your key, your decision. The tool does not need to ration anything because there is nothing for the tool to ration.

The reframe

The AI provider relationship is shaped like a utility relationship, not a subscription relationship. You have a meter. The meter belongs to you. lognote is the appliance that plugs into it.

Pick the provider that fits the meeting. For the calls that don’t need the lift, reach for the local path. Reach for a frontier key when you do. Don’t let a tool make that choice for you, on this tool or any other one.