Security and privacy

What leaves your computer depends entirely on which providers you pick in Settings. Below is the complete list of connections, files and permissions — so your IT department or data protection officer can check it rather than take our word for it.

This page describes the Sidecue desktop app for Windows. Last updated: September 2026.

Need the same on two pages, to forward on? Download the compliance pack (PDF) — who plays which role, a checklist of data processing agreements per provider, and wording for informing participants.

The short version

There is no Sidecue server that receives your conversations. The app has no user account, syncs nothing and collects no telemetry. Audio and transcript go only to the provider you choose, using your own API key.
The fully local mode removes even that. Transcription through whisper.cpp and the language model through Ollama run on your own machine — nothing from the conversation goes out at all.
No bot joins the meeting. Sidecue reads the audio already playing on your computer, so no extra guest appears in the meeting. It does not notify other participants by itself — tell them you use an AI assistant wherever your law or the meeting's rules require it.
The meeting is not recorded. Audio is processed as it streams and never lands on disk as an audio file.

Where data goes

This is every outbound connection the app makes. The first six depend on your choice in Settings — pick a different provider and the connection is never established.

Function	Endpoint	What is sent	When
Cues — Claude	`api.anthropic.com`	A slice of the transcript and conversation context	When the selected provider is Anthropic
Cues — OpenAI	`api.openai.com`	A slice of the transcript and conversation context	When the selected provider is OpenAI
Cues — Gemini	`generativelanguage.googleapis.com`	A slice of the transcript and conversation context	When the selected provider is Google
Cues — local	`localhost:11434`	Never leaves the machine	When Ollama is selected
Transcription — Deepgram	`wss://api.deepgram.com`	The audio stream	When Deepgram is selected
Transcription — Whisper API	`api.openai.com`	Audio chunks, one per utterance	When Whisper over API is selected
Transcription — local	a process on your machine	Never leaves the machine	When whisper.cpp is selected
Licence	`sidecue.io` (our licence server)	Licence key and computer name — nothing else	On activation, then a check once every 24 h
Updates	`dl.sidecuehq.com`	A version-number lookup	Handled by the Windows update mechanism

What is not on this list: a Sidecue server receiving conversation content, telemetry, usage analytics, crash reporting, user accounts, or cross-device sync. The list is complete — there is no "other" row.

That list is about the application. The website is a separate matter and deserves an honest description: sidecue.io counts visits. It does so with Umami — a counter we run ourselves, on our own server, in the same cluster as this page. What it stores is the page address, country, device and browser type, where the visit came from, and clicks on the download and pricing buttons. There are no cookies, no identifier that survives a day, no way to reconstruct an individual from that data, and none of it reaches a third party — which is why this site asks for no cookie consent. The application sends nothing to that counter; it knows about website traffic and nothing else.

The site also carries a form for the release announcement. If you leave your address there, we store: the email address, the date, which language version of the page you were on, which pricing shelf the form sat next to, the wording of the consent you were shown, and the IP address the submission came from — the last two as evidence of consent. It all lives in a database on our own server, not in a third-party mailing tool. Right after you sign up we send a message with a confirmation link; when you use it, we also store the time and IP address of the confirmation. The messages themselves are sent by Brevo (France) as a processor — it receives the address only at the moment of sending and keeps no copy of our list. Apart from the confirmation we will send one message, on release day, and use your address for nothing else. To have it removed, write to [email protected] — we do it right away and without asking why.

The AI and transcription endpoints are those providers' public APIs. Sidecue does not proxy that traffic: you connect directly, with your own key, and settle your bill with them without us in the middle. The data-processing terms that apply are therefore theirs — worth reading their terms and data processing agreement before your first confidential call.

The fully local mode

For conversations that must not go to the cloud, Sidecue can be configured so that no content whatsoever leaves the machine:

Transcription: whisper.cpp — a speech model running on device, no key needed.
Cues: Ollama at http://localhost:11434 — no key, no per-token cost.
Chat reading: always local anyway — through the Windows accessibility layer or OCR on the window you point it at.

The data path is then: audio → whisper.cpp → context → Ollama → the panel on your screen. It works with no internet connection at all.

One exception you should know about. The licence check contacts our own server at sidecue.io once a day and sends only the licence key and the computer name — never conversation content. The app keeps working for 7 days without a connection, so an air-gapped machine is workable but needs to be reconnected periodically. If you need a licence that runs entirely offline, write to [email protected].

How to verify it yourself

You do not have to trust us on this. With local mode configured, run a network monitor (Windows Resource Monitor, or a firewall that logs connections), hold a conversation, and check whether Sidecue.exe opens any connection beyond localhost. The strictest test is a firewall rule blocking the app's outbound traffic entirely — in local mode Sidecue runs normally with it in place.

What is written to disk

What	Where	Notes
Settings	`%APPDATA%\Sidecue\settings.json`	Provider, device and preference choices. No API keys.
API keys and licence	Windows Credential Manager	Entries named `Sidecue:*`, protected by your Windows account. They never enter the settings file.
Diagnostic log	`%APPDATA%\Sidecue\sidecue.log`	Application flow. Conversation text is recorded as a length only (e.g. `[128 zn.]`, a character count) unless you deliberately turn on diagnostic mode. The file rotates at 1 MB.
Offline licence	`%APPDATA%\Sidecue\licencja-offline.txt`	A signed file you install yourself, if you use an offline licence. Readable — you can open it in Notepad. Checked on this computer only, with no network connection.
Problem report	`%TEMP%\sidecue-zgloszenie-*.zip`	Created only when you click “Report a problem”. Contains the log and the diagnostic card. Nothing is uploaded — the file stays on your disk until you send it yourself.
Conversation transcript	not written	Lives in memory for the duration of the session.
Audio recording	not written	Audio never reaches the disk, in any mode. In local mode the utterance is piped to the whisper.cpp process on standard input, with no intermediate file.
Meeting summary	wherever you save it	A file exists only if you click save and choose a location.

Uninstalling the app does not remove %APPDATA%\Sidecue or the Credential Manager entries — delete them by hand if you are passing the computer on.

Keeping your notes off a shared screen

When you share your screen with a client, your private notes and cues should not be shared with it. That is why every window the app opens is marked with the system flag WDA_EXCLUDEFROMCAPTURE. In practice the cue panel does not appear in a shared screen or in a screenshot — not even when you share the whole desktop. On your own monitor it looks perfectly normal.

The flag covers all Sidecue windows, not just the panel — Settings and the summary too. The panel shows the status itself: while the protection is active, a green line confirms your notes are kept off shared screens.

The limits of this protection. The flag works at the Windows level and defends against software capturing the screen. It does not defend against a phone pointed at your monitor, or somebody standing behind you. The app can also be launched deliberately with the --allow-capture switch, which turns the protection off — it exists for recording tutorial material and never engages on its own.

Permissions the app asks for

Microphone and system audio — so it can hear the conversation at all. You can mute the microphone at any time from the panel or with Ctrl + Alt + M; muted audio is not passed on anywhere.
Reading window content — through the Windows accessibility layer (the same one screen readers use) or OCR, and only for the window you point it at. Sidecue does not scan your desktop or other applications.
Installation without administrator rights — the MSIX package installs into your user profile and does not touch system directories.
Window capture without the yellow border (graphicsCaptureWithoutBorder) — OCR mode only. Windows draws a yellow outline around a window being captured; that outline appears on your screen alone and never reaches the other side of the call, so all it does is distract the person using the app. Turning it off requires your consent — Windows asks once and you may refuse, in which case the border stays. The permission does not widen what is read: still only the window you point the app at.
Permissions the app never asks for: camera, location, contacts, calendar, mailbox. Sidecue does not integrate with your calendar and needs no access to a Microsoft or Google account.

Phone view — a port on your local network

The app has an optional feature that mirrors the cues to a phone on the same Wi-Fi. It is off by default. Once enabled it opens port 8750 on the local network and requires a one-time token in the address; Windows will ask for a firewall exception, and the private network profile is enough. The traffic never leaves your network, but if company policy does not allow this, simply leave the feature off.

Questions we get asked

Can anyone at Verilabs see the content of my meetings?

No. There is no infrastructure to receive it — the content never reaches an address of ours. Even if we wanted to hand something over, there would be nothing to hand: all that reaches us is the fact that a licence key was activated.

Will my conversations be used to train AI models?

Not by us — we do not have them. In cloud mode this is governed by your agreement with the provider you chose, and this is an area where the details surprise people. As of September 2026; worth checking at the source before your first confidential call:

Google Gemini — for users in the European Economic Area, Switzerland and the UK Google does not use prompts or responses to improve its products, even on the free tier. Outside those countries the free API tier does, and the content may be read by a human reviewer; only a project with active Cloud Billing excludes that (Gemini API terms, version of 28 April 2026).
Deepgram — by default may use recordings to evaluate and improve its models, including with third-party speech recognition providers. An option in Sidecue's Settings switches that off by adding mip_opt_out=true to every connection — at a higher rate, according to sources outside Deepgram — and so does a signed data processing agreement (compliance and privacy).
Anthropic — the data processing addendum covers paid commercial products including the API, and excludes training on customer data.
OpenAI — API data is not used to train models unless you opt in (how your data is used).

This is the easiest thing to miss. After installation Sidecue suggests Gemini and Deepgram because they are the quickest to get running — not because they are the most careful. In Europe a free Gemini key does not hand your transcript over to improve Google's models, but Deepgram may use recordings until you turn on the opt-out in Settings or sign an agreement with it. For conversations under professional privilege, do one of the two — or use the fully local mode.

If you would rather not settle that question at all, use local mode. It disappears, because nothing from the conversation leaves your computer.

Will participants see Sidecue when I share my screen?

No. The app's windows are excluded from screen capture at the Windows level — the details and the limits of that protection are described above.

Does the app record the meeting?

It writes no recording file. Audio is processed as it arrives and is gone once the text has been recognised. The transcript is not written to disk either — the only thing that persists is what you save yourself.

Where are my API keys kept?

In Windows Credential Manager, under entries named Sidecue:*, protected by your Windows account. They are not in the settings file, not in the log, and are never sent anywhere but to the provider they belong to.

Does the app work without an internet connection?

In local mode, yes. The licence check needs a connection once a day, with 7 days of slack — after that a session will not start until the app can refresh the licence. For a machine that stays off the network permanently that is not enough; write to [email protected] in that case.

Can I use Sidecue for conversations covered by professional privilege?

Technically, local mode is built for exactly this case: content never leaves the computer, so no third party is entrusted with the data. Whether your profession and situation permit using a tool that processes the conversation — and whether the other party must be told — is your assessment to make as the data controller. We are happy to answer the technical questions that assessment needs.

Can I get a data processing agreement?

In local mode there is nothing to entrust — no data reaches us. In cloud mode the processing is entrusted to the AI provider you selected, and that is who you sign with. If your legal team needs written confirmation of the scope of processing on our side, write to [email protected] and we will prepare it.

How do I report a vulnerability?

Write to [email protected] with a description and steps to reproduce. We treat those reports first and will not pursue claims against anyone conducting security research in good faith.

A question that isn't here?

Write to [email protected]. If it comes from an IT department or a data protection officer, you will get specifics back — including what the app cannot do.

See also the getting started guide, which walks through choosing providers step by step, including the fully local setup.