See exactly where your information goes.
The questions an IT team asks are easier to answer with a picture than a paragraph. Each diagram below has the plain answer above it and the specifics underneath, in a drawer you can open if you want them. The written question-and-answer version lives on our Connecting securely page.
What connects to what
Every connection into MetriFi starts on your side, and you can switch any of them off. Nothing runs the other way: there is no route from MetriFi into your core, your online banking, or your member records, because none was ever built.
The red box is your own systems. MetriFi cannot reach it. A visitor gets there by clicking a link on your website, the same way they would from anywhere else on the internet.
The technical detail
- Transport
- HTTPS over TLS 1.2 and TLS 1.3 only; TLS 1.0 and 1.1 are refused. Cloudflare terminates TLS and provides DDoS protection and edge filtering in front of the application.
- How your staff sign in
- Server-side sessions with Secure, HttpOnly, SameSite cookies, expiring after 120 minutes of inactivity. Passwords are a minimum of 12 characters, hashed with bcrypt at a work factor of 12, and are never stored or recoverable in plaintext.
- How an AI assistant is authorized
- OAuth 2.0 authorization code flow with PKCE, alongside the device authorization flow, dynamic client registration, and standards-based discovery with JWKS. Access tokens expire in one hour and refresh tokens in 30 days. Every connection can be listed and revoked from your MetriFi account, and revoking takes effect on that assistant's next request.
- Google Analytics
- Read-only, aggregate site metrics, from a connection you authorize yourself in your own browser. MetriFi never receives your Google password. The resulting token is encrypted with AES-256-CBC and HMAC-SHA-256 authentication; the encryption key is held in the server environment, separately from the database, and never in source control.
- OpenAI
- Used to generate research prompts, keyword terms, and conversion recommendations. Most of that work runs on background job queues, and none of it runs on your website, so a slow or unavailable provider cannot slow your live site. Under MetriFi's master services agreement, neither party may submit nonpublic personal information to any AI system or model provider.
- Where it runs
- The application, its database, and its backups run in US regions. Your website's source is a private repository on GitHub, and the published site is served from Vercel's edge network. Two services are globally distributed on purpose — Cloudflare, and that edge network — because both exist to serve public page content from the location nearest each visitor.
- What is not there
- Nothing is installed on your network, no customer-side server is required, and no inbound connection to your systems exists. A link from a published page to online banking or a loan application is an ordinary hyperlink, followed by the visitor's browser. No credential is shared and no server-to-server call is made.
What happens before a change goes live
Nothing an AI drafts reaches your live site on its own. Two of these ten steps are people: someone has to open the preview and approve it, and their role has to allow publishing. The assistant proposes; your team decides.
The numbers are the reading order. The middle row runs right to left, which is why every box carries its step number.
The technical detail
- Where a draft lives
- On its own branch in your site's repository, never on the branch that production is built from. Publishing is what moves it across. Until then the live site is byte-for-byte unchanged.
- The preview
- A private URL behind a per-site password, so an unpublished change is never publicly readable and never indexed. A client review link carries its own opaque token; that link reaches its own review and nothing else, and it is worth treating like any shared document link.
- The disclosure gate
- A publish is refused outright if a page requires a disclosure that has no approved text. Two further checks — pairing an advertised rate with its disclosure, and elevated review of protected pages — run today as warnings shown to the person publishing rather than as hard blocks, and are being moved to enforcement as the fleet is brought up to them.
- Approved wording
- Every disclosure is stored once, as your institution's own approved language. Pages reference it by name rather than pasting its words, so the exact text appears everywhere it is used and cannot drift. The AI never writes or edits disclosure language.
- Attribution and rollback
- Every change records what changed, when, and who or what made it; a change made through an assistant is stamped with the person who authorized it. The version in production is tagged before a publish, so rolling back is republishing a known-good version rather than reconstructing one.
- Scheduled rate changes
- Rates can be set to publish themselves at a future time. The scheduler rebuilds the site at that moment; it does not change anything that was not already approved.
Who gets in, and how you shut it off
A person and an AI assistant get in different ways, then pass exactly the same checks. Being an assistant grants nothing beyond what the person who authorized it already has, and one click in your account ends it.
Every fact in these four bands is one of our written answers to a vendor security questionnaire, not a summary of one.
The technical detail
- Passwords
- A 12-character minimum, enforced by one shared rule across registration, reset and change. Hashed with bcrypt at work factor 12. Reset tokens are single-use, time-limited and rate-limited, and the reset flow does not reveal whether an email address is registered.
- Rate limits
- Five login attempts per minute per client, six password-reset requests per minute, and per-route limits between 30 and 120 requests per minute on public endpoints. A failed login returns a generic message, so the login form cannot be used to find out which accounts exist.
- Keeping clients apart
- Every tenant-owned record carries a team identifier or belongs to a parent record that does, and membership of the owning team is re-verified on every request. The core records are also filtered to the authorized team automatically at the database layer, and a record belonging to a parent object is re-checked against that parent's team. Automated tests that assert one team reads none of another's data run on every change to the platform.
- Programmatic access
- Tokens carry product-specific scopes, and a token with no product scope is rejected at the endpoint. Administrative scopes are reserved for MetriFi staff, stripped at issuance for customer tokens, and not advertised in public API discovery.
- Roles
- Four team roles — Owner, Admin, Editor, Contributor — with least privilege applied by default. Only Owners and Admins can invite, remove, or change the role of a team member. The full table is on our Connecting securely page.
- MetriFi staff
- Named personnel holding a platform administrative role can access client data for support. That access is granted by platform role, not by membership in your team, and is limited to people with a legitimate business need at the minimum level their responsibilities require.
- Not available today
- Federated single sign-on from your own identity provider, such as Google Workspace or Microsoft Entra ID, is not currently scheduled for implementation.
What a site visitor actually touches
Your website is a set of files. It runs no database, no content management system, no plugins, and no login-based admin screen, which removes the categories of vulnerability behind most website compromises. Everything a member types happens somewhere else.
The purple boxes are your own vendors. What a member enters there is collected and held by them, under their controls, and never reaches MetriFi.
The technical detail
- What the site is
- Pages are pre-built into finished HTML and served from the location closest to each visitor. There is no server software assembling a page on each visit, no database to wait on, and no administrative login to attack.
- Vendor integrations
- Online loan and account applications are links to your third-party tools. Embedded integrations, such as online banking, use the code your vendor provides. Neither is different from a traditional build.
- Analytics
- Page-view events go to your own Google Analytics 4 property. MetriFi reads aggregate metrics from that property if you connect it, and never writes to it.
- Security scanning
- The site is scanned before launch and rescanned quarterly, with the results reported to you. Website source sits in version-controlled repositories covered by the same continuous dependency, secrets and code scanning as the platform itself.
- Availability and recovery
- Served from a globally distributed edge network with no single server to fail. There is no database to corrupt, every version is preserved, and a bad change is undone by republishing the previous version. The complete source of your site is portable, so it can be republished elsewhere if it ever needs to be.