By Henry Ho, Founder of Geode
Founder’s Notes #1
When we were designing Geode, the hardest product decision was not whether local AI mattered.
It did.
The harder question was whether Geode should stop there.
A pure local product is easier to explain. Everything stays on the device. No cloud. No ambiguity.
A cloud-first product is easier to scale. Upload the audio, run the best available models, return the transcript and summary. One workflow. One architecture. One business model.
Geode chose a harder path:
Local by default. Cloud only when you choose it.
That decision came from two problems I kept seeing in audio AI.
The first is privacy.
Recorded audio is not ordinary app data. A transcript is already sensitive, but the original audio often carries more: tone, hesitation, and context that may never appear in the written version.
As cloud AI becomes more capable, it also needs more context to be useful. That is the tension. The better the model gets, the more it wants to know.
For some recordings, that trade-off is acceptable. For others, it is not.
The second problem is cost.
Cloud AI has a meter, even when the product tries to hide it. Long recordings consume compute. Summaries consume tokens. Better models cost more to run.
For someone who transcribes a short meeting once in a while, this may not matter. For someone recording interviews, lectures, or long client conversations every week, it matters quickly.
At some point, the user stops thinking about the work and starts thinking about the meter.
I do not think serious audio work should feel that way.
So Geode starts locally.
You can record and transcribe without sending audio to a cloud model by default. On Mac, Geode also supports heavier local work such as speaker separation and on-device summaries, because the Mac has the hardware for more serious local processing.
That choice created a real trade-off.
We did not make every feature identical on every device. iPhone and iPad are better for capture and mobile transcription. Mac is better for deeper local processing. A cloud-first product could hide that difference by sending everything to servers.
We chose not to hide it.
But Geode is not local only.
There are recordings where cloud AI is useful. A user may want a stronger model for a specific file, a faster result, or simply decide the recording is not sensitive.
I do not want the product to make that decision for them.
This is where many audio AI tools still make the wrong trade-off.
Cloud transcription tools give users powerful models, but no real choice about where the work happens. Local transcription tools give users control, but often make cloud AI external, manual, or disconnected from the workflow.
Geode’s bet is that the next generation of personal AI tools should not force that choice.
The product should have a default. The user should have the final say.
That is why Geode is local by default.
Not because cloud AI is bad.
Because recorded audio belongs to the person who created it, and that person should decide when it is worth leaving the device.