Why Cue gives you both on-device and cloud AI

Why does Cue offer both on-device and cloud AI? Because local-first is practical before it is ideological.

The everyday path can be quicker because it avoids a network round trip. It can keep working offline once the local model is downloaded, and the durable result stays in files you control. When a task benefits from a more capable connected model, that option is available too. Cue lets you choose without making the cloud the only place your work can live.

The short answer

Local-first means the local path is useful on its own. Local-only means connected processing is not offered at all.

Cue gives you both because different work deserves different boundaries. Use on-device transcription when you want a path without a network round trip, when audio should stay on your Mac, or when you need to work offline. Keep the result as local Markdown. Choose a cloud model or connected agent for a specific job when it offers capability the local path does not.

That is a choice, not a blanket performance promise. Speed and output quality still depend on the model, Mac, language, network, and task. The important questions are what crosses the network, when it crosses, and who decides.

Question	Local-first answer	Connected option
Where does speech become text?	On the device when an on-device mode is selected	Audio is sent to the chosen transcription service
Where does the useful result live?	In a readable local file or local app state	A service may hold a secondary copy if the product supports sync
Can the core work continue without a connection?	Yes, for the parts backed by local models and files	Connected features wait for the network
Can outside AI use the work?	Not unless you choose a supported connection	Selected text or context is sent for a specific job
Is the boundary visible?	The active mode and destination should be clear	The user confirms before processing leaves the device

Check four boundaries, not one privacy label

“Local” can describe several different things.

Processing: where audio becomes text.
Storage: where the transcript, note, or meeting output is kept.
Connection: what is sent to another model or service later.
Action: what an agent or integration is allowed to read or change.

These boundaries can differ in the same workflow. Speech may be transcribed on the Mac, saved as a local note, then copied into a cloud coding agent. The first two steps were local. The last one was not.

That does not make the earlier steps fake. It means “local-first” is not a magic word for the entire chain.

The original local-first software paper makes the same distinction at the architecture level. The local copy is primary, while servers can still help with access and synchronization. Local-first tries to keep the user's work useful without making a remote service the only keeper of it.

One real workflow

Imagine you need to capture a rough product decision on your Mac.

Check the active transcription mode. Choose an on-device model when the audio should stay on the Mac.
Speak the thought. Capture the decision, the reason, and the unresolved question.
Keep the result locally. Save the reviewed note as a readable file you can inspect and move.
Connect only when the next job needs it. If you want a cloud model to critique the decision, select the relevant passage, remove anything unnecessary, and send it deliberately.
Keep the source. The local note remains useful even if the connected service is unavailable later.

This is different from a cloud-first workflow in which the remote account is the only place the note can be read or changed. It is also different from local-only software, which would reject the connected critique entirely.

The point is not to make one choice for every task. It is to make each boundary understandable.

When cloud processing helps

Connected models and services can be useful when a task needs a capability that is not available on your Mac, when you want to work with another person, or when you deliberately want an outside agent to use selected context.

That value comes with a different boundary. Audio, text, or context must leave the device. The exact handling then depends on the service and its current terms.

A good local-first product should make that change visible before it happens. A small mode label can be more useful than a large privacy slogan.

The cloud is not the villain. Hidden dependence is the problem.

What local-first does not promise

Local-first is a design choice, not a complete security policy.

It does not automatically mean:

Every feature works offline.
No data ever leaves the device.
Local models are always faster or more accurate.
Local files are backed up.
Another app cannot read a file you gave it permission to open.
A connected provider keeps nothing after processing.
The software is open source or can be self-hosted.

Offline, on-device, private, encrypted, user-owned, and local-first are related ideas. They are not synonyms.

For sensitive or regulated work, check the whole path: microphone, transcription mode, file location, optional correction, connected tools, backups, and retention terms. A local label cannot answer those questions for you.

Why Cue offers both local and connected options

Cue is local-first because the local path is useful on its own, not because Cue bans the cloud.

Choose on-device transcription when you want audio to stay on your Mac or need dictation without an internet connection. Choose cloud transcription when you deliberately want a connected provider. Cue shows which transcription mode is active, so the boundary is visible before you speak.

The same idea continues after capture. Cue Notes keeps the source material as local Markdown you can inspect, move, and keep. If a subscription changes access to hosted processing, it should not decide whether you can still open your own notes. The durable work remains useful as files outside one app.

If you later want an outside model to critique a passage or a compatible agent to use selected context, that is a new connected step. Cue MCP makes that connection optional and permissioned.

A real workflow can use both paths:

Capture a sensitive product thought with on-device transcription.
Keep the reviewed source in a local Markdown note.
Select only the passage that needs outside help.
Send that passage to a connected model deliberately.
Review the response while the original local note remains intact.

The cloud step does not erase the local-first design. It does create a different boundary, and the product should make that change obvious.

That is why Cue offers both options. The goal is not to make “local” or “cloud” win every time. The goal is to let you choose the right boundary for the work in front of you.

Read Cue's current privacy policy for data-handling terms. Check it again when those terms matter to a specific piece of work.

If you only need occasional built-in text entry, macOS Dictation may be enough. Apple tells users to check Keyboard settings to see whether general Dictation is processed on-device or sent to Siri servers. A dedicated tool such as Cue Dictation makes sense when voice has become a regular input method and you want an explicit local or cloud choice.

When this approach is not enough

Choose a different tool or deployment when your work requires controls Cue does not currently provide.

That may include a browser-first collaborative editor, a managed enterprise environment, a fully air-gapped workflow, or a formally reviewed data-processing agreement. Local files and on-device transcription are useful properties, but they do not replace organizational security and compliance requirements.

Local processing can also be a poor fit when the required model is too large for the available Mac, has not been downloaded, or does not support the language or task well enough. In those cases, a connected service may be the practical choice if its boundary is acceptable.

Frequently asked questions

Does local-first mean no internet?

No. A local-first product keeps the primary work useful on the device. It may still use the internet for sync, collaboration, updates, or optional AI processing.

Is local-first the same as offline?

No. Offline describes whether a particular task works without a connection. Local-first describes the product's relationship to the user's data and local copy more broadly.

Is local-first automatically private?

No. Local processing can reduce one kind of exposure, but privacy also depends on permissions, storage, backups, telemetry, connected tools, and what the user sends later.

Can a local-first app use cloud AI?

Yes. The cloud use should be optional, visible, and limited to the task the user chose. The local copy should remain useful without that service.

Does on-device transcription mean nothing ever leaves my Mac?

It means the transcription step can remain on the Mac when that mode is active. A later correction, AI request, sync service, or agent connection may cross a separate boundary.

Can Cue dictate without an internet connection?

Cue offers on-device transcription. Once the required local model is available on the Mac, that transcription path does not require an internet connection. Connected models and services still do.

Sources

Originally accessed August 22, 2026. Cue product and policy pages rechecked September 9, 2026.

Choose the boundary yourself

Download Cue for Mac when you want one product that gives you an on-device path for capture, local Markdown for the work you keep, and connected processing when you deliberately choose it.

Start with Free voice-to-text for Mac for setup, or read why voice should work like an input layer.