Skip to main content
Puntego runs cleanly under a strict Content Security Policy. The boot script and runtime are nonce-aware: you pass your per-request nonce once, and Puntego reuses it for every script and style element it injects. No 'unsafe-inline' and no 'unsafe-eval' are required for the guide to load, chat, point, and run Verified Actions. This page gives you a baseline policy that covers the core experience, then the exact additions to make when you enable Verified Actions or voice. Treat the samples as a starting point and confirm them against your own application’s directives — most sites already serve a CSP, and the values below are meant to be merged into it rather than to replace it.

Minimal baseline policy

This is the smallest policy that loads the guide, lets it chat, and lets it point and highlight on your page. Replace {NONCE} with your per-request nonce value (the same one you put on the boot script). A real header is a single line; it is wrapped here for readability.

What each origin is for

Test before you ship. The injected stylesheet carries your nonce, so a nonce-based style-src admits it without 'unsafe-inline'. We do not promise a nonce-only style-src is sufficient for every setup — your other app directives and any third-party styles still apply — so verify the policy in a staging environment and watch the browser console for CSP violations before rolling it to production.

Add Verified Actions (Turnstile)

Verified Actions protect writes with a Cloudflare Turnstile challenge. Turnstile loads a script and renders its challenge in a frame, so add its origin to two directives:

Add voice

Voice is optional per workspace. When it is enabled, the runtime streams the visitor’s speech to the Puntego worker, which transcribes it, and the guide’s spoken replies back, both over WebSockets to your worker origin, so extend connect-src:
wss://worker.puntego.com is your data-gp-api origin with the wss: scheme. A CSP host source with https: does not match wss: (CSP3 scheme matching only upgrades http: to https: and ws: to wss:), so list it explicitly. Without it the guide still speaks, by requesting each sentence over HTTPS with a slower start, and listening falls back to other paths; the browser logs a CSP violation for each refused socket. The player that plays the stream is an AudioWorklet loaded from a blob: URL, like the microphone capture: a policy that already lets voice capture run needs nothing more for it. No browser connects to the speech-recognition provider any more: the worker holds that connection. An older policy that still lists wss://stt-rt.soniox.com keeps working; you can remove it. OpenAI realtime voice no longer needs a CSP entry: the browser sends its connection offer to the Puntego worker (already in connect-src), which starts the call with OpenAI, and the audio then flows over WebRTC, which connect-src does not govern. An older policy that still lists https://api.openai.com keeps working; you can remove it. A strict-CSP site must also grant microphone access through a Permissions-Policy header, or the browser blocks capture before Puntego ever asks:
If voice provider credentials are not configured in production, voice remains unavailable and the provider is not contacted; text guidance continues. The internal mock provider is limited to development and tests.

Complete policy (all features)

With Verified Actions and voice both enabled, the merged policy looks like this:
Trusted Types: puntego-loader is used by the npm/programmatic loader, puntego-boot by boot.js, and puntego by runtime HTML and feature-script sinks. A Trusted-Types-enforced page must allow the names for the integration paths it uses.

Install snippet

Use the standard snippet when you do not serve a nonce-based CSP. Keep crossorigin="anonymous" in both forms.

Standard

Strict CSP (with nonce)

Pass the same nonce value to both nonce and data-nonce:
Copy the canonical snippet from Dashboard → Install so your app ID stays aligned with the deployed worker and embed assets.

How nonce propagation works

You only set the nonce once. The boot script reads it from either nonce or data-nonce, then the runtime reuses that value for everything it adds to the page:
  • The boot script reads your nonce from nonce or data-nonce.
  • The runtime stamps the same nonce on every follow-on script it loads (the runtime chunk and, when voice is active, the voice chunk).
  • The runtime stamps the same nonce on the single stylesheet it injects for the guide’s shadow root.
Because Puntego propagates the nonce instead of inlining unguarded scripts or styles, your script-src and style-src can stay nonce-based with no 'unsafe-inline'.

Sub-apps in iframes

If the guide works inside one of your sub-apps in an iframe (Sub-apps in iframes), the sub-app’s own pages load the child script, and the child reads Puntego’s public keys to check the pass the guide brings for each connection:
That key request carries no cookie and nothing about your visitors; everything else the child says goes only to its parent page. Have the sub-app send frame-ancestors with your site’s origin so only your page can embed it, and keep your page’s frame-src (or child-src) to the sub-app origins you allowed:

Next steps

  • Install Puntego — the full script tag, configuration attributes, and domain allowlist.
  • Framework guides — read a per-request nonce from next/headers and pass it through the @puntego/react provider, plus single-page-app routing and a no-code Google Tag Manager template.
  • Verified Actions — what the Turnstile origins protect and how approved actions are configured.
  • Security — how Puntego isolates the guide, redacts PII, and keeps telemetry clean.