Prywatność

# NETWORK.md — every connection Whisperer can make

Whisperer is a **local-first** application. Your audio, recordings, transcripts, and summaries are
processed on your machine and stored only in your configured storage folder. This document is an
exhaustive, behavior-verifiable inventory of **every** network connection the app is capable of
making — nothing else happens. You do not have to take our word for it: run a network monitor
(Little Snitch on macOS, `lsof -i` / `ss -tunp` on Linux, a firewall log on Windows) and confirm the
app opens only the connections listed here.

*This file is kept in sync with the code. If you find an outbound connection not listed here, that is
a bug — please report it.*

## The complete list

| # | Host | When | What is sent | Contains your data? | Avoidable? |
|---|------|------|--------------|---------------------|------------|
| 1 | **The LLM endpoint you configure** (`llm.endpoint_url`; **no default** — until you set one, summaries are off and nothing is sent) | When a meeting summary is generated, and a `/models` health probe when you test the connection | The meeting **transcript text** (for summarization) or a models list request (for the probe) | **Yes — transcript text.** It goes only to the endpoint *you* set. Out of the box no endpoint is set, so nothing is sent at all; point it at a server on your own machine and nothing leaves it. If you point it at a cloud API, your transcript goes there **by your choice** | Yes — disable summaries / leave the endpoint unset. Meetings still record and transcribe fully locally |
| 2 | **huggingface.co** (and its model CDN) | First time a model is needed: the Whisper-family speech-to-text models, and — only if you opt in to the Parakeet engine in Settings — the Parakeet-TDT-0.6B-v3 weights from the Hub repos `mlx-community/parakeet-tdt-0.6b-v3` (macOS) or `istupakov/parakeet-tdt-0.6b-v3-onnx` plus the `istupakov/silero-vad-onnx` segmenter it needs for long audio (Linux/Windows), each pinned to a fixed revision. Diarization models come from here **only if you added a Hugging Face token** (optional fallback — the default path is row 5) | A standard model **download** request; for the diarization models, your Hugging Face **token** in an `Authorization` header (only if you provide one) | No — these are downloads *to* your machine. Your token authenticates the download; no audio/transcript is uploaded | Partly — once models are cached locally they are not re-downloaded. Diarization is optional |
| 3 | **huggingface.co/api** (token check) | Only when you click "Test" on the optional Hugging Face token field in Settings ▸ Diarization | Your HF token in an `Authorization` header, to verify it can access the diarization model | No | Yes — it only runs when you test the token |
| 4 | **saykeep.app** (the update appcast) | At most once/day, only if you enabled updates (you are asked once at first run) | A GET of a static XML file. The request carries a standard User-Agent naming the app and its version — `Saykeep/<version> Sparkle/2.9.6` — and nothing else: no machine id, no account, no licence key, no cookie, no identifier that could single you out or link two requests together. Downloads and installs the release only when you consent | Your Saykeep version, and that a copy of Saykeep is checking. Nothing else about you or your machine | Turn it off in Settings ▸ General ▸ Check for updates and no request is ever made |
| 5 | **models.saykeep.app** (the speaker-label model mirror) | Once, only if speaker labels (diarization) are enabled and the model is not yet on disk — shown in the "Getting Saykeep ready" download window | An anonymous GET of a static ~30 MB tarball (`diarization-v1.tar.gz`) — no machine id, account, token, or any identifier (the request carries only a generic HTTP-client name). The file is checksum-verified before it is installed | Nothing about you or your machine is sent | Yes — turn speaker labels off (Settings ▸ Diarization) and no request is ever made; once installed it is never re-downloaded |

*Row 2 note (2026-09-05): 0.4.0 also made one small metadata request to huggingface.co each time
a cached Whisper model loaded (at launch, when the meeting model loaded, and on every switch
between the two) — the loader was handed a Hub repo name instead of the on-disk path. Found and
proved on the wire while fact-checking the website
(`docs/ceo/evidence/2026-09-04-hf-launch-metadata-request.md`), disclosed here the same day, and
**fixed in 0.4.1**: a cached model now loads from its local snapshot with no request (wire proof:
three requests → zero across load → meeting → dictation). It carried no token and no identifier.
If you are still on 0.4.0, update, or block huggingface.co — the model loads from disk either way.*

*Row 4 note: this check is off by default and never opens a socket while off.*

*Row 4 honesty note (corrected 2026-08-30): this row previously claimed the check sends "no version
string". That was wrong — the update engine (Sparkle) sends its standard User-Agent, which names the
app version, and nothing in the app overrides it. The row now states what actually goes over the
wire. Verify it yourself with Little Snitch or `tcpdump` per the "watch it live" section below —
that is the point of this document.*

*Row 5 note: this is the token-free replacement for the Hugging Face diarization download (rows 2–3). Those two rows now apply only if you chose to add a token yourself.*

That is the entire list.

## What Whisperer **never** does

- **No telemetry, analytics, crash reporting, or usage tracking.** No analytics SDK is linked into
  the app — and you can hold us to it from the outside: the rows above are everything, and the wire
  is checkable.
- **No accounts, no login, no cloud sync.** There is no Whisperer server. Nothing is uploaded.
- **Your recordings, transcripts, and summaries are never sent anywhere** except the transcript text
  to the LLM endpoint *you* configured (row 1). They are stored only in your storage folder.
- **No license phone-home.** License keys are verified **offline** against a compiled-in public
  key — no activation server, no network call, ever.
- **The diagnostics export** (Settings → Advanced) writes a **local file** you choose whether to
  share; it makes no network call, and it redacts secrets.

## Local helpers are not network connections

Whisperer runs some local subprocesses — `pw-record` (Linux system audio), the bundled
`whisperer-syscapture` (macOS system audio), `osascript`/`caffeinate` (macOS), and your OS's
file-reveal command. These are **on-device** and open no network sockets.

## How to verify

1. **Watch it live:** macOS → Little Snitch; Linux → `sudo ss -tunp | grep whisperer` or `lsof -i -a -p <pid>`; Windows → Resource Monitor / firewall log. With summaries pointed at a local endpoint and models already cached, the only connection you will see during normal use is the **update check** — at most one request a day to the Saykeep appcast, and only if you said yes to it at first run (row 4). You will see your Saykeep version in that request's User-Agent, and nothing else identifying. The one other connection you may see is a **single** download of the speaker-label model the first time you open Saykeep with speaker labels enabled (row 5); it never repeats once the model is on disk. Turn the update check off in Settings and — once that one-time model download has happened, or with speaker labels off — you will see **zero** outbound connections.
2. **Inspect your config:** Settings → Advanced → "Export diagnostics…" shows your resolved
   `llm.endpoint_url` (credentials redacted) so you can confirm where summaries would go.
3. **Hold us to the contract:** Whisperer is closed-source, so instead of "read the code," this file
   *is* the auditable surface — the only things that ever open a socket are the rows above. If you
   ever observe a connection that is not listed here, that is a bug; report it and we will treat it
   as one.