Quickstart

Send your first PII-redacted LLM request through the Treza AI Control Plane. You'll create an API key, configure a proxy, point your OpenAI client at Treza, and check the audit log.

Prerequisites

• A Treza account — sign in at trezalabs.com (email sign-in; the Starter plan is free)

• An API key from your LLM provider (e.g. an OpenAI sk-... key)

1. Create a Treza API key

In the dashboard, go to API Keys and create a key with the redact:proxy scope. Add redact:run and redact:log if you also want the standalone redaction and log endpoints.

The key (treza_live_...) is shown once at creation. Store it in a secrets manager — you'll send it as the Authorization: Bearer header on every request.

2. Create a proxy

Go to Control Plane in the dashboard and create a proxy:

1. Choose the upstream provider (OpenAI on Starter; Azure OpenAI, Anthropic, or a custom endpoint on Pro+).

2. Paste your upstream provider key. It is envelope-encrypted with KMS at rest, never returned by any API, and used server-side — so your application never needs to handle the provider key again.

3. Copy the proxy id (proxy_...).

Don't want Treza to store your provider key? Leave it off the proxy and pass it per-request in the x-model-key header instead.

3. Point your client at Treza

The proxy endpoint is OpenAI-compatible, so existing SDKs work with a base-URL change.

curl

curl https://trezalabs.com/api/redact/chat/completions \
  -H "Authorization: Bearer $TREZA_API_KEY" \
  -H "x-treza-proxy: proxy_1718039482_ab12cd34e" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o-mini",
    "messages": [
      { "role": "user", "content": "Draft a reply to [e-mail] confirming her SSN 123-45-6789 was updated." }
    ]
  }'

The upstream provider receives the redacted prompt:

Draft a reply to [EMAIL_1] confirming her SSN [SSN_1] was updated.

TypeScript (OpenAI SDK)

import OpenAI from 'openai';

const client = new OpenAI({
  apiKey: process.env.TREZA_API_KEY,                  // treza_live_...
  baseURL: 'https://trezalabs.com/api/redact',    // SDK appends /chat/completions
  defaultHeaders: { 'x-treza-proxy': 'proxy_1718039482_ab12cd34e' },
});

const completion = await client.chat.completions.create({
  model: 'gpt-4o-mini',
  messages: [
    { role: 'user', content: 'Draft a reply to [e-mail] confirming her SSN 123-45-6789 was updated.' },
  ],
});

console.log(completion.choices[0].message.content);

Python (OpenAI SDK)

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["TREZA_API_KEY"],                 # treza_live_...
    base_url="https://trezalabs.com/api/redact",     # SDK appends /chat/completions
    default_headers={"x-treza-proxy": "proxy_1718039482_ab12cd34e"},
)

completion = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[
        {"role": "user", "content": "Draft a reply to [e-mail] confirming her SSN 123-45-6789 was updated."},
    ],
)

print(completion.choices[0].message.content)

Streaming is supported. Set stream: true and the proxy forwards the upstream provider's Server-Sent Events stream straight through, so token-by-token output works with the standard OpenAI SDKs. Treza never rewrites the response body — placeholders are rehydrated client-side from the x-treza-rehydration map, which is sent as a response header before the first token. Because a placeholder can be split across SSE chunks, rehydrate against the accumulated text rather than individual deltas.

4. Inspect the response

Every proxied response carries Treza headers:

Header	Meaning
x-treza-request-id	Unique id for this redaction request (matches the audit log)
x-treza-mode	standard (managed software redaction) or tee (Enterprise enclave)
x-treza-rehydration	Placeholder → original map (only when requested, see below)

If the model's reply references placeholders, request the rehydration map and restore values client-side:

curl -i https://trezalabs.com/api/redact/chat/completions \
  -H "Authorization: Bearer $TREZA_API_KEY" \
  -H "x-treza-proxy: proxy_1718039482_ab12cd34e" \
  -H "x-treza-rehydrate: 1" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"Reply to [e-mail]"}]}'

# Response header:
# x-treza-rehydration: {"[EMAIL_1]":"[e-mail]"}

const { data, response } = await client.chat.completions
  .create(
    { model: 'gpt-4o-mini', messages },
    { headers: { 'x-treza-rehydrate': '1' } },
  )
  .withResponse();

const map: Record<string, string> = JSON.parse(
  response.headers.get('x-treza-rehydration') ?? '{}',
);
let text = data.choices[0].message.content ?? '';
for (const [placeholder, original] of Object.entries(map)) {
  text = text.split(placeholder).join(original);
}

5. Try redaction without forwarding

To see what the pipeline detects — without calling any upstream and without using your request quota — use POST /api/redact/run (requires the redact:run scope):

curl https://trezalabs.com/api/redact/run \
  -H "Authorization: Bearer $TREZA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "Call John Smith at [phone] about card 4111 1111 1111 1111"}'

{
  "redacted": "Call John Smith at [PHONE_1] about card [CC_1]",
  "entities": [
    { "type": "PHONE", "placeholder": "[PHONE_1]", "start": 19, "end": 28 },
    { "type": "CC", "placeholder": "[CC_1]", "start": 40, "end": 46 }
  ],
  "requestId": "req_lx2k9a_1f3b7c2d",
  "mode": "standard",
  "modelVersion": "...",
  "recognizerVersion": "..."
}

6. View the audit log

Each request (proxied or standalone) writes an audit entry scoped to the API key that made it. Fetch entries with the redact:log scope:

curl "https://trezalabs.com/api/redact/log?limit=20" \
  -H "Authorization: Bearer $TREZA_API_KEY"

{
  "entries": [
    {
      "ts": "2026-06-11T18:21:09.412Z",
      "requestId": "req_lx2k9a_1f3b7c2d",
      "source": "proxy",
      "entityCountsByType": { "EMAIL": 1, "SSN": 1 },
      "attestationRef": "shared-software-redaction",
      "modelVersion": "...",
      "recognizerVersion": "..."
    }
  ]
}

Query parameters: limit (1–500, default 20) and since (ISO timestamp). Audit entries contain entity counts by type — never the redacted values themselves. Request counts and entities-redacted insights are also available in the dashboard under Usage & insights.

Next steps

• Proxies & Policies — multiple providers, pausing proxies, custom policies

• Usage & Billing — usage insights, audit export, and how requests are counted

• Plans & Pricing — included volumes and what happens when you hit your limit

• Attestation — prove redaction ran inside a Treza Enclave (Enterprise)