Deliverd Screen

Deliverd Screen

A pre-send check for every prompt. Find and redact personal data, secrets and your own confidential terms before text goes to an AI model.

• Local. No account, no key, no network call. The text never leaves your process, unless you turn on the optional Microsoft Purview check, which sends it to your own Microsoft tenant and nowhere else.
• Reversible. Values become placeholders such as [EMAIL_1] (the same value gets the same placeholder throughout), and restore() puts them back in the model's answer.
• Bring your own model for names. Patterns can't find people's names, but a model can. Plug in a Hugging Face token classifier, Presidio, or any model with the same output.
• No dependencies. MIT.

npm install @deliverd/screen

import { screen } from "@deliverd/screen";

const checked = await screen("Reply to mary.byrne@example.ie about IBAN IE29AIBK93115212345678"); checked.verdict; // "redact" checked.text; // "Reply to [EMAIL_1] about IBAN [IBAN_1]"

const answer = await model.generate(checked.text); checked.restore(answer); // the real values, back in the answer

What it finds

Email addresses, phone numbers, IBANs, payment card numbers (Luhn-checked), UK National Insurance numbers, Irish PPS numbers (check character), US Social Security numbers, IP addresses (v4 and v6), dates of birth, API keys and tokens (AWS, GitHub, Slack, Stripe, Anthropic, OpenAI, Google, JWTs and more), private keys, passwords in connection strings and .env files, words that mark special-category data such as health or religion, and your own terms.

The four answers

By default the model may receive internal data and no personal data. Pass allow: { personalData: true, classification: "confidential" } for a model you have approved for more.

Names: plug in a model

import { pipeline } from "@huggingface/transformers";

import { tokenClassifier, presidio } from "@deliverd/screen/adapters";

const ner = await pipeline("token-classification", "Xenova/bert-base-NER"); const checked = await screen(prompt, { detectors: [tokenClassifier((t) => ner(t))] });

// or a Presidio analyzer you run yourself: // docker run -p 5002:3000 mcr.microsoft.com/presidio-analyzer await screen(prompt, { detectors: [presidio({ url: "http://localhost:5002" })] });

A detector is a function from the text to { kind, start, end } spans, so any model works, including Perplexity's PII-Tracer and GLiNER. Check a model's licence before you ship it. Run it where the text already is: a detector that sends the text to a hosted service has already sent it.

Vercel AI SDK

import { generateText, wrapLanguageModel } from "ai";

import { screenMiddleware } from "@deliverd/screen/ai-sdk";

const model = wrapLanguageModel({ model: openai("gpt-5"), middleware: screenMiddleware() }); const { text } = await generateText({ model, prompt });

The middleware screens the whole conversation at once, so one value keeps one placeholder across messages. It restores the answer, including a streamed one where a placeholder arrives split across chunks. On a deny or an ask it throws ScreenBlockedError and the model is never called. It works with AI SDK 7 by default; pass specificationVersion: "v3" for AI SDK 6.

CLI, pre-commit and CI

cat prompt.txt | npx @deliverd/screen --redact   # the redacted text

npx @deliverd/screen --json notes.txt # kinds and offsets, never values npx @deliverd/screen --fail evals/*.jsonl # exit 1 if anything is found npx @deliverd/screen --terms clients.txt doc.md # your own words, one per line

In GitHub Actions, uses: deliverd-dev/deliverd-screen@v0 with paths: runs the same check; for pre-commit, the hook id is deliverd-screen. Both are in deliverd-dev/deliverd-screen.

MCP

claude mcp add deliverd-screen -- npx -y @deliverd/screen mcp

This gives an agent in Claude Code, Cursor or Codex one tool, screen_text. It returns the verdict and the redacted text, and never the values it replaced, so they don't enter the agent's context through the tool.

Microsoft Purview

If your organisation writes its data loss prevention rules in Microsoft Purview, the same rules can apply here, for any model and any tool. Pass a Purview check and the stricter answer wins: a Purview block denies, and a warn is an ask.

import { entraAppToken, purviewCheck, screen } from "@deliverd/screen";

const purview = purviewCheck({ token: entraAppToken({ tenantId, clientId, clientSecret }), userId: entraObjectIdOfThePerson, app: { name: "Claims assistant", entraAppId: clientId }, }); const checked = await screen(prompt, { purview });

This is the one thing in the package that uses the network, and only when you pass it: the text goes to Microsoft Graph (processContent), in your own tenant, and nowhere else. The Entra app needs Content.Process.All, and the DLP rule must be created with New-DlpComplianceRule (Purview does not apply a portal-only rule to an Entra-registered app).

For the MCP server, set DELIVERD_PURVIEW_TENANT_ID, DELIVERD_PURVIEW_CLIENT_ID, DELIVERD_PURVIEW_CLIENT_SECRET and DELIVERD_PURVIEW_USER_ID (and optionally DELIVERD_PURVIEW_ON_UNAVAILABLE: allow, ask or deny).

A net, not a certificate

Structured identifiers are found well. Names are found only as well as the model you plug in. Redacted text about a person is still personal data under GDPR. Use this to send fewer identifiers, never as proof that a prompt was clean.

When a person has to decide

A library can only tell you to stop. Deliverd's Data Gate makes the call. deliverd.screen() in @deliverd/sdk runs these same detectors on your machine and sends Deliverd only the counts ("two email addresses, one IBAN"). The verdict comes back from your organisation's policy and from what its AI register approves each model to receive. On an ask it puts the question to the right person in Slack, Teams or on their phone, and every check is recorded. It takes the same detectors.

Licence

MIT. The detectors are shared with Deliverd's SDKs and held to one set of test vectors, packages/conformance/screening.json.