Silvertongue writes in a famous voice, or in your own, when you ask it to, and keeps as little as it can to do that. This page covers the browser extensions, Silvertongue and Silvertongue Self, the try box on each voice page, the Telegram and Discord bots, and the MCP server. Last changed 4 November 2026.
What the extension reads
On the sites it supports, the S button's script notices which box you are writing in and reads nothing else until you press the S or the shortcut. Then it reads the box and the conversation that box belongs to, as already shown on the page: the post you are replying to and the posts above it, or the subject and newest messages of an email thread. It never reads the replies below yours, your feed, your inbox, other conversations, other tabs, or password, payment, search and code fields. It never scrolls, expands or fetches anything to read more.
What is sent, and when
Only when you press Reply, Rewrite or Write: what it read (unless you left it out), your text or the part you selected, the voice, the length, the site's name and your writing settings. They go to api.silvertongue.dev (earlier versions use its old address, silvertongue.auscaster.com), which sends them to the model that writes the draft, currently Google's Gemini API, and streams the draft back to you. Nothing is sent while you type.
When you set up your own voice, the writing you paste is sent once, when you press Study, to the same model, which studies how you write.
On a voice page, the box sends the point you typed and the voice, only when you press Write it, the same way. Your browser keeps a token for the box, as an extension install does, and it gives the box three drafts.
On Telegram and Discord, the bot receives only what you send it: the voice and your point, when you pick a voice or use /as. It never reads a group's or a server's messages. Your Telegram or Discord user ID is kept only as a keyed digest that cannot be turned back into the ID, which is how your three drafts are counted; your name, username and the chat you are in are not kept.
What is kept
- Your text and the draft: held in the service's memory for 15 minutes, so a retried request gets the same draft back, and never written to a disk or a database. The record of each draft keeps only fingerprints (digests) of the request and the draft, which is how fills are counted.
- Your install: a random install ID made when you install, and a token only your browser holds. The service keeps the token's digest, never the token.
- Your account, if you sign in: the number Google gives your Google account, never your name or email address, when you signed in, and a key only your browser holds for this install, kept as its digest. Your fills are counted against the account, and your install's usage counts are joined to it. Signing out ends the key.
- Usage counts, which never carry text: the install ID, the voice, whether it was a reply, a rewrite or a new post, the length, how it ended, its latency and cost, and where, as a supported site's name such as X or Gmail, or the kind of page such as a forum. Never a domain or an address. A searched name is kept only as its slug, such as keanu-reeves, and only when it is a name. Usage counts are kept for 90 days, then only as daily totals with no install in them.
- Your own voice, if you set one up: a profile of how you write (descriptions, counts of habits such as how you use capitals and apostrophes, and phrases you use) and at most sixteen short lines it quotes from your writing, kept against your account until you delete it. The writing you pasted is held in the service's memory only while it is studied, and for 15 minutes after so a retried study gets the same result, and never written to a disk or a database.
- Visits to these pages: counted once a visit, by the kind of page and where you came from: the tag on a link we posted, or the kind of site that linked here, such as Google or TikTok. The tab remembers the source until you close it, so the visit counts once.
- Site analytics: these pages use Google Analytics, Microsoft Clarity and Cloudflare Web Analytics, as most sites do, to see how pages are used. Google and Microsoft set their own cookies; their privacy policies cover what they keep.
- Fault reports: when a site's box or conversation cannot be read, which step failed and which version of the site's recipe was used. Never any text.
- The access log: the server that runs the service and the MCP server keeps a standard web server log of each request's address, time and path for 14 days, to run and protect the service. Nothing else keeps an address: rate limits count requests per address in memory only.
Your controls
Settings turn usage counts and fault reports off, each on its own. The record of each draft continues, because fills are counted from it. You can hide the S on a site or everywhere, and stop it reading the conversation on any site. You can delete your own voice on its profile page at any time. Removing the extension removes its token with it; the browser then opens a page that counts the removal.
What Silvertongue does not do
It does not sell anything about you, show ads, or train models on what you write.
The MCP server
The server at mcp.silvertongue.dev only reads: it searches the voices and returns their profiles and packets, and your own model writes the draft. It keeps nothing about your requests beyond the server's access log. It counts requests per address in memory to limit them.
Who handles data for Silvertongue
Amazon Web Services runs the service and its database. Google signs you in, and Google's Gemini API writes the drafts. Stripe will take payments when plans open.
Questions and deletion
Ask through auscaster.com/studio. Usage counts tied to an install expire after 90 days; ask, and they are deleted sooner.