Docs
tacet runs on your machine. Your client sends to it, it swaps each listed value for a stand-in, sends the request to your provider, and puts the real values back in the answer.
Draft a refund e-mail to Maria. She paid with card 4539 1488 0343 6467 and wants the money on IBAN NO93 8601 1117 947. Her address is maria.k@fjordmail.no and her phone is +47 412 55 309.
Draft a refund e-mail to Maria. She paid with card 4242 4297 1375 7590 and wants the money on IBAN NO09 1609 9990 717. Her address is rixil.t@rixv.example and her phone is +47 460 65 764.
Hi Maria, the charge on 4539 1488 0343 6467 is cancelled. The refund goes to NO93 8601 1117 947 within 3 working days. We will text +47 412 55 309 when it lands, and the receipt goes to maria.k@fjordmail.no.
Try it in 2 minutes
1. Start it. It listens on http://127.0.0.1:8787/v1 and sends to openrouter.ai/api/v1 unless you pass --upstream.
2. Point your client at it. Keep your provider key: it is passed through, never stored.
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "http://127.0.0.1:8787/v1",
apiKey: process.env.OPENROUTER_API_KEY,
});3. Send a request. Start it with --show to print what the provider sees for each one.
What is hidden
The formats on this list never leave your machine. A value in any other format goes out as written.
| Format | How it is found | Stand-in |
|---|---|---|
| Service keys | About 25 known key prefixes: OpenAI, Anthropic, OpenRouter, AWS, GitHub, GitLab, Slack, Stripe, Google and more | Same prefix, length and character classes |
| Secrets in config | .env, YAML, JSON and code values with a secret name, passwords in connection URLs, PEM private keys, JWTs, base64 that wraps a secret | Same shape, header kept |
| Wallet keys | EVM 0x + 64 hex, Solana base58 or 64-number keypair | Another key of the same format |
| Seed phrases | 12 to 24 BIP-39 words with a valid checksum, or after seed, mnemonic, recovery | A valid phrase of the same length |
| Card numbers | 13 to 19 digits, known card prefix, Luhn check | A test card number, valid Luhn, same grouping |
| IBAN | Country length table and mod-97 | Same country and length, valid check digits |
| E-mail addresses | Address syntax, Unicode allowed | A rare word at example.com, .net or .org |
| Phone numbers | International with +, North American, French, UK formats | A fiction range where one exists, else other digits |
| Your own terms | A file of words you add, whole words, any case | Rare pseudo-words, case kept |
Not hidden on purpose: dates, cities, organisations, public wallet addresses, git SHAs, UUIDs, lockfile hashes, IP addresses.
Config
Optional. Flags on the command line, or a file at ~/.tacet/config.json.
| Flag | Default | |
|---|---|---|
| --port, -p | 8787 | Local port, bound to 127.0.0.1 only. |
| --upstream, -u | openrouter.ai/api/v1 | Any OpenAI-compatible base URL. |
| --terms, -t | none | A file of your own terms, one per line. |
| --show, -s | off | Print what the provider sees for each request. |
| --config, -c | ~/.tacet/config.json | port, upstream, terms, toolPolicy, disabled. |
| --key-file | ~/.tacet/key | Your local key, made on first run. |
toolPolicy says which values may go back into the arguments of a tool call. By default only e-mails, phones and your terms do. Secrets stay as stand-ins.
{
"port": 8787,
"upstream": "https://openrouter.ai/api/v1",
"terms": ["Project Larch", "Ada Vance"],
"toolPolicy": {
"write_env": ["ENV_SECRET", "API_KEY"],
"*": ["EMAIL", "PHONE", "TERM"]
}
}What works today
- The formats on the list are found and swapped on your machine, with no model and no network call.
- The same value gets the same stand-in in every turn, with no stored map.
- Every part of the request is covered: system, user, assistant, tool results, tool arguments, tool schemas.
- Real values come back while the answer streams, also if the model turns a seed phrase into a numbered list.
- Secrets are not put back into tool-call arguments, unless you allow it for one tool.
- A request with a part it does not know is refused, and nothing is sent.
What does not work yet
- Names and street addresses are not found unless you put them in your terms file.
- Formats not on the list go out as written: unknown key formats, a password in a sentence, IP addresses, national ID numbers.
- A secret split across lines or string pieces is not found.
- URL-encoded values and e-mails written as name [at] domain are not found.
- If the model rewrites a stand-in in a new way, that value stays as the stand-in in the answer.
- Images, audio, files and embeddings are refused, not masked.