T·OLMACH
Postern Labs · interpreter & scribe
MIC LIVE — CONTINUOUS CAPTURE
I speak
They speak

Consecutive interpretation — each utterance lands ~1–3 s behind the speaker. This is not simultaneous interpreting, and we won’t pretend it is.

    Nothing interpreted yet. Hold a side and speak — or type below (the fully local input path).
    The source transcript always stays visible so you can spot mis-hearings.

    ● no server · no account · key stays in this browser translator: · $0.0000 ·
    elapsed 00:00 ≈ $0.0000 · 0 in / 0 out tok

    Speaker — tap to tag what’s said next

    Speaker labels are manual or content-inferred (inferred) — the browser cannot identify voices.

    Steno — raw transcript (stays in this tab)
    Structured minutes

    Minutes will build here — updated automatically every ~2 minutes while the meeting runs, and on demand.

    Settings

    Stored only in this browser’s localStorage and sent only to api.anthropic.com — Tolmach has no server that could ever see it. Get a key at console.anthropic.com.

    Both billed to your own Anthropic account at your account’s rates. The cost readout in Minutes mode counts the actual streamed tokens.

    Speech-to-text engine
    Translation engine

    Tolmach auto-picks a voice matching the target language; your preference wins when its language matches. Voices marked “⚠ network voice” are synthesized by the OS vendor’s servers, not on-device.

    The vault also locks when this page is hidden or closed. The decryption key exists only in this tab’s memory while unlocked.

    Erase everything

    Wipes your API key, all settings, this session’s transcripts and minutes, and the encrypted vault from this browser. Nothing else exists to erase — there is no server copy. Exported vault files on your disk are yours to delete.

    Privacy — who sees what

    Tolmach is a static page. There is no Postern server, no account, no telemetry, and no analytics. Your API key and settings live in this browser’s localStorage; transcripts and minutes live in this tab’s memory unless you save them to the encrypted vault. The honest ledger:

    ◮

    Your voice → your browser’s vendor (whenever the mic is on)

    Voice input uses the browser’s Web Speech API. In Chrome and most browsers this sends your audio to the browser vendor’s servers for transcription, under that vendor’s terms — so speech input is not fully private. In conversation mode the mic is push-to-talk only. In Listen and Minutes modes the microphone is continuously hot for as long as the mode runs — the red MIC LIVE banner stays up the whole time, and stopping the mode kills the recognizer immediately. Want zero audio leaving the machine? Type instead — typed input is fully local until you send it to Claude. The real fix is the on-device Whisper engine slot, wired but not bundled here.

    ◮

    Transcript & minutes text → Anthropic (when a segment is sent)

    Each translation segment and each minutes update is sent from your browser directly to api.anthropic.com, authenticated with your key — no middleman relays or logs it. Anthropic processes it under your agreement with them (their usage and data-retention terms apply). If you never send, nothing is ever transmitted.

    ●

    Spoken translations — on-device

    Text-to-speech uses your operating system’s voices via speechSynthesis; audio is generated locally and does not leave the device. Exception: some OS voices are network voices — the voice picker marks those “⚠ network voice” so you can avoid them.

    ●

    No speaker identification happens anywhere

    The browser has no speaker diarization. Speaker labels in the minutes are either your manual taps or the model’s content inference, and every inferred label is marked (inferred). Tolmach never records or stores audio — transcripts only; less data is the default.

    ●

    API key & settings — this browser only

    localStorage of this browser, sent only to api.anthropic.com. “Erase everything” in Settings wipes all of it instantly.

    ●

    The vault — encrypted, local, and honestly limited

    Saved sessions are encrypted with AES-256-GCM in this browser’s IndexedDB; the key is derived from a passphrase we never see and never store (PBKDF2-SHA-256, 600,000 iterations — weaker than Argon2id against offline GPU attacks on weak passphrases, so pick a strong one). Anyone holding the exported vault file or the disk sees ciphertext only. Honest limits: a browser origin is not a hardware vault — whoever controls this browser profile while the vault is unlocked (malware, a malicious extension, someone at your unlocked machine) can read notes and capture the passphrase as you type it. A forgotten passphrase means the data is unrecoverable by design. The real-product fix is Postern Vault’s TEE key-wrap.

    —

    Postern sees nothing

    Static page, no server, no account, no telemetry, no analytics. The only network destinations in this page’s code: api.anthropic.com, plus the browser’s own speech service when voice input is used. One readable HTML file — verify it yourself.

    What we won’t claim: Tolmach is not “100% private.” Anything you send to a model necessarily reaches Anthropic, and spoken input may reach your browser vendor. What Tolmach guarantees is no extra parties: no Postern servers, no resellers, no analytics. And it won’t call itself a simultaneous interpreter either — it is consecutive, ~1–3 s behind the speaker.

    Vault — encrypted archive

    checking…