SDKs

The drop-in proxy keeps PII away from the LLM provider. The SDKs go further: they tokenize messages in the user's browser, so your backend, database, and logs hold only tokens, and reveal them again when the reply reaches the user.

Preview

The SDKs are in preview and not yet published to npm or PyPI. Package names and APIs may change before 1.0.

Who sees plaintext

Pick a tier by where you want plaintext to stop. Each tier includes the protection of the ones before it.

Tier	How	LLM provider	Your backend	Your frontend
Proxy	Point your server's LLM client at NoPII	Tokens	Plaintext	Plaintext
Tier 1	The browser SDK reroutes provider calls made from the page	Tokens	Not involved	Plaintext
Tier 2	The browser SDK tokenizes requests to your own routes	Tokens	Tokens	Plaintext

No tier takes your frontend out of scope. Your page's JavaScript runs on the same origin as the text the user types, so it can read it. Hosted input elements that close that gap are planned.

Choose a package

Package	Runs in	Use it to
@nopii/browser	Browser	Tokenize requests to your routes, reroute provider calls, and reveal replies
@nopii/react	React apps	Wire the browser SDK into React and the Vercel AI SDK's useChat
@nopii/node	Node.js servers	Mint client sessions, forward tokens to the LLM through the proxy, tokenize for indexing
nopii-sdk	Python servers	The same as the Node.js SDK, for sync and async Python
@nopii/mock-server	Your machine	Develop and test offline against an in-memory NoPII

How Tier 2 fits together

A typical chat app uses one SDK on each side.

text

Browser (@nopii/browser)                 Your server (@nopii/node or nopii-sdk)         NoPII proxy
------------------------                 ---------------------------------------         -----------
1. Asks your server for a client session
                                         2. Mints one with its secret key
3. Tokenizes the message with NoPII
4. POST /api/chat, tokens only       ->  5. Forwards the tokens, tokens-only mode    ->  6. Explains the tokens to
                                                                                            the model, calls it
7. Reveals the streamed reply        <-  8. Streams the tokenized reply back         <-
   for the user

Your server never handles plaintext in this flow. It stores tokens, logs tokens, and passes tokens to the model, which is told what each kind of token stands for so it can still reason about them.

Credentials

Credential	Where it lives	Used for
`nsk_...`	Your server only	Minting client sessions and reveal grants, server-side tokenization
`Client session`	The browser or app, for up to an hour	Tokenizing and revealing as one of your end users. Recommended.
`npk_...`	Embedded in a web page	Development and low-risk apps. Only works from the origins you list.

Create keys under API Keys in the admin console. See Keys and reveal policy for what each can do.

Fail closed

When NoPII cannot be reached, the SDKs throw rather than send the request unprotected. A protected route always fails closed. Only provider interception can be told to fall back, with failOpen: true, and it should stay off, because it sends plaintext exactly when protection is broken.

Tokens reach your backend by design

In Tier 2, your servers receive values like [NAME: 8f2kQ1xZ]. Store them as they are. To show a user their own history later, return the tokens and let the SDK reveal them in the browser.

Browser SDK: start here for a web app
Keys and reveal policy: who may turn a token back into plaintext
Client API reference: the endpoints the SDKs call