Security and limits
Public site keys, allowed sites and apps, test and live keys, the limits in numbers, and the CSP lines for your site.
Site keys are public
A site key (mwg_pk_live_… or mwg_pk_test_…) ships inside your page or app, so it isn't a
secret. What protects it is the list of allowed sites and apps you give it:
- Websites are checked by the browser's
Originheader, which a web page can't fake. An entry is an exact origin (https://shop.example.com,http://example.com:8080) orhttps://*.example.comfor any subdomain at any depth (neverexample.comitself, and a wildcard takes no port). - Apps send their iOS bundle ID or Android package name (
X-Mawgood-App: ios:com.example.app). An honest app can't use another app's key, but a modified app can fake its ID. App protection reduces misuse; it doesn't prevent it entirely. The limits below cap what such misuse can spend. - The list stops other websites from using your key, because browsers set
Originthemselves. It doesn't stop a script running outside a browser: any client can send anyOriginheader. The limits below cap what such a script can spend. - Up to 20 sites and apps per key. Each call is checked on every request, and the browser only
gets CORS answers for the origin it was checked against, plus the
origin_not_allowedrefusal itself when the key is valid (it names only the origin to add).
Change key makes a new key with the same sites and apps; the old one keeps working for 7 days, so you can update your site or app (End the old key now stops it early). Revoke key stops a key at once on every site and app, for a key that leaked. Open bubbles on a revoked or expired key end.
The two secrets never ship: your public API key (mwg_sk_…) and the
identity secret (mw_sk_…) stay on your server.
Test and live keys
| Key | Works from |
|---|---|
mwg_pk_test_ | The sites and apps you list, plus localhost, 127.0.0.1 and [::1] on any port, always. |
mwg_pk_live_ | Only the sites and apps you list. Loopback addresses never work, and can't be added to the list. |
Both talk to the same Mawgood: a test key's chats are real chats, answered by the Mawgood and charged like any other (1 credit per AI reply). Use a test key on your machine and a live key on your real site or app.
Limits
| What | Limit | Over it |
|---|---|---|
| Opening a session | 30 a minute per IP | 429 rate_limited, with Retry-After |
| Every chat API call | 600 a minute per IP | 429 rate_limited |
| New chats per device | 10 a day | 409 new_chats_limit |
| New chats per IP, per Mawgood | 200 a day | 409 new_chats_limit |
| New chats per Mawgood | 2,000 a day (you can lower it) | 409 chat_volume_limit |
| Messages per chat | 20 a minute | 429 rate_limited |
| Uploads per visitor | 30 every 10 minutes | 429 rate_limited |
| "Typing" signals per chat | 30 a minute | 429 rate_limited |
| AI replies per Mawgood on this channel | 1,000 an hour (you can lower it) | The chat goes to your team, and the bubble says so |
| Live connections | 3 per visitor, 300 per IP per Mawgood | Refused; the SDKs wait and retry |
| A file | 10 MB: JPEG, PNG, WebP, GIF images, PDF, and voice notes | 413 file_too_large, 415 unsupported_type |
| A message | 4,096 characters, up to 10 files | 400 validation_failed |
Many real customers in Egypt share one mobile IP (carrier NAT), which is why the per-IP limits are high. Images and voice notes your visitors send are read by AI (payment receipts, transcription) within an hourly budget per Mawgood the same size as the reply cap; past it they are kept for your team, unread.
The SDKs handle all of this for you: they back off on 429, keep messages that failed in their
outbox, and show «try again later» on the new-chat limits.
Content-Security-Policy
If your site sends a Content-Security-Policy header, the web bubble needs
these sources (production hosts; apps don't use CSP):
script-src https://cdn.mawgood.ai;
connect-src https://mawgood.ai wss://rt.mawgood.ai https://*.r2.cloudflarestorage.com;
img-src https://mawgood.ai https://*.r2.cloudflarestorage.com blob: data:;
media-src https://*.r2.cloudflarestorage.com blob:;
style-src 'unsafe-inline';
font-src https://your-font-host.example;| Directive | Why |
|---|---|
script-src | The script tag, https://cdn.mawgood.ai/chat/v1.js. Not needed when you install from npm (the code is in your bundle). |
connect-src | The chat API and its Server-Sent Events fallback (https://mawgood.ai), the realtime socket (wss://rt.mawgood.ai), and file uploads, which go straight to Mawgood's storage (Cloudflare R2). |
img-src | Your bubble's images (the launcher image, the Mawgood's photo), files in the chat (short-lived links to R2), and previews of files the visitor picks (blob:). Catalog card photos come from wherever your catalog's images are hosted: add those hosts too, or use https:. |
media-src | Voice notes and voice replies (R2), and the visitor's own recording (blob:). |
style-src | The bubble adds its styles in its own Shadow DOM with a <style> element. |
font-src | Only when the look sets shape.font.webFontUrl: the host of that font file. With the default system font, leave it out. |
About the R2 host. Files aren't served from a public bucket: uploads and file links are
short-lived signed URLs on Mawgood's storage account's S3 endpoint,
https://<account-id>.r2.cloudflarestorage.com. Its exact host isn't published here, so the lines
above use the wildcard. To narrow it, copy the host from any file link in your bubble (it's the same
for every file) and use it instead of *.r2.cloudflarestorage.com.
For staging, use staging.mawgood.ai and wss://rt-staging.mawgood.ai instead.
Privacy
The bubble shows a one-line privacy notice, linked to the privacy page set on the card, before the
first message. The device id lives in first-party storage only (localStorage on web, secure
storage in apps): no cookies, no tracking. A device isn't a contact until its customer writes.
Deleting a contact deletes its chats.