Privacy

Draft, not yet in force. This policy will be reviewed and finalised before run sharing is released. Items in [brackets] are still to be confirmed.

The short version

If you use Tiny Agent without turning on sharing, we receive nothing from the app.
This website uses no cookies, analytics or third-party scripts.
If you opt in, we receive only the runs you choose to share, and you can delete them at any time.

Who we are

NinaCorp [legal entity name, address]. Contact: [privacy contact email].

The app, without sharing

The local model, your notes, tasks and saved conversation history stay on your computer. Features that reach the internet contact their service directly and never NinaCorp. Optional Voice & ideas has two connections: Realtime sends microphone audio, typed messages and the current canvas to OpenAI using separate API billing. Spark with local speech processes audio on-device and sends the current thought, canvas and last four saved text turns through Codex using your ChatGPT sign-in and Spark allowance. Explicit local Whisper setup downloads packages from PyPI and model data from Hugging Face; transcription runs offline. Microphone tests use local speech inference, not Spark or the paid API, and do not save test words or audio. There is no automatic paid fallback. TinyAgent does not record audio; it saves up to 40 recent text turns, idea cards and Spark run diagnostics locally. An API key entered in the connection panel stays in the background process for that launch, not in the canvas. Web search contacts the configured search provider.

Live sketch uses local tiny speech and loopback Ollama, without Spark or OpenAI calls. Normal edit instructions, bounded prompts, replies, errors, timing, working HTML/CSS, sketch snapshots and the agreed scope record (goal, audience, included and parked ideas, any open question) are saved locally. Audio and provisional transcripts are not recorded. Its matching tiny-speech check does not save test words or alter the page. Closing a chat is soft deletion; scratch files and diagnostic history remain recoverable on the device. Preview content has no network assets or executable scripts, and replacing the editor draft requires explicit review.

Optional Phone Chat shares the selected Chat title and up to 30 bounded recent text turns with a phone you approve locally. It uses private Tailscale HTTPS, a one-use two-minute QR and revocable eight-hour sessions. Messages still run through the PC's configured providers and existing approval gates. It sends nothing to NinaCorp. The phone page uses no analytics or localStorage; browsers can still retain screenshots or other copies. This is separate from the future research-sharing service, and is not an independently audited security product.

This website

Our web server doesn't keep access logs. Your IP address is held in memory for up to an hour to limit abuse of the sharing endpoints, and isn't written to disk.

Shared runs (opt-in)

What: the contents listed on the Contribute page, plus a random install ID so you can export or delete your runs. IP addresses are used only briefly to limit abuse and aren't stored with shared runs.
Why: to study and improve how the harness understands requests, and to publish aggregate research findings.
Legal basis: your consent, which you can withdraw at any time.
Where: our server hosted by IONOS [data-centre region].
Who else: AI model providers used for analysis [provider names], under terms that don't let them train on the data.
How long: until you delete it, or [24 months] after it was shared, whichever comes first.

Your rights

You can access, export and delete your shared runs from within the app. Depending on where you live, you may also have rights to rectification, restriction and objection, and to complain to your data-protection authority. Contact [privacy contact email].