There is nothing here to collect.

No accounts, so no account data

There is nothing to sign up for on this site. We hold no name, no email address, no password and no profile, because there is no mechanism here that could create one.

No trackers, of any kind

No Google Analytics, no tag manager, no advertising or remarketing pixel, no session recording, no heatmap, no embedded video, no social widget, no chat bubble. Not even a font CDN: the two typefaces are served from our own server.

No cookies

This site sets none. It keeps three display preferences in your browser’s own local storage, which never travels to a server. That is why there is no consent banner — there is nothing to consent to.

Masked logs, thirty days

The web server writes one line per request, with your IP address truncated before it is written and any cookie header discarded. The file rolls off after thirty days.

Your Mantle data never reaches us

The product is self-hosted. There is no telemetry, no crash reporting, no usage statistics, no licence check phoning home, and no hosted tier for any of it to land in.

Cross Works Engineering (Pty) Ltd publishes Mantle and operates this website at mantle-ai.tech. For the narrow set of data described on this page, we are the responsible party under South Africa’s POPIA and the controller under the GDPR.

Privacy questions, requests and complaints: licensing@crossworks.engineering.

The site runs on a single virtual private server that we rent and administer ourselves. No content delivery network, tag provider or analytics vendor sits in front of it, so the only parties in a position to observe a request are our hosting provider and us.

mantle-ai.tech is a static export. Every page you can reach was rendered to plain HTML, CSS and JavaScript at build time and copied to the server. There is no application server, no database, and no session. Nothing you do here is written anywhere except the access log in the next section.

That is not a privacy feature we added. It is how the site was built, and the privacy follows from it: a server with no database cannot accumulate a profile of you even if someone later decided it should.

The site has no way to know who you are, and no place to write it down if it did.

How the masking works. IPv4 addresses are truncated to a /24 and IPv6 addresses to a /32, so 203.0.113.47 is written as 203.0.113.0. Both address fields the server produces are masked, not just the obvious one — masking only the first leaves the full address sitting in the second, which is a mistake worth naming because it is easy to make. The truncation happens before the line reaches the disk, so the complete address is never at rest anywhere.

What we do with it. Occasionally we generate an aggregate traffic report from those masked lines — visits per page, broad geography, which crawlers are hammering us. The report is built on the server, read by us, and never published or handed to a third party. There is no advertising or attribution use, because there is no advertising.

Why there is no banner. We consider a truncated address with no cookie, no identifier and no account to attach it to non-identifying, and the retention is short and fixed. Consent tooling here would collect more about you than the logging it was asking permission for.

All three are strictly functional, hold no identifier, and are readable and deletable by you at any time: clearing site data for this domain removes them, and the site works exactly the same afterwards. Nothing else is written to your browser — no local database, no service worker cache of your activity, no fingerprinting.

• —No analytics or product-analytics script
• —No advertising, remarketing or conversion pixels
• —No session recording or heatmaps
• —No A/B testing or personalisation
• —No third-party fonts or CDNs — the typefaces are self-hosted
• —No embedded video, maps or social widgets
• —No newsletter, CRM, lead capture or contact form
• —No live-chat widget
• —No comment system
• —No fingerprinting of any kind

Two files at the root of the domain do exist and are worth explaining rather than leaving for someone to find: /BingSiteAuth.xml and /taaft.txt. Each holds a fixed verification string proving we control this domain, to a search engine and to a directory we listed the product on respectively. They are static text. They run nothing, and they collect nothing.

Links out to GitHub, the documentation and the model providers we name are ordinary links. Following one puts you on someone else’s site under their privacy policy, not ours.

It is shared and public

Everyone who opens the demo lands in the same environment. Treat everything in it as visible to anyone else who opens it, because it is.

You are signed in automatically

The server attaches a demo session to requests on their way through, so there is nothing to register for and no credential of yours is involved anywhere.

It is reset without notice

The brain is seeded from a fixture. Anything that accumulates in the environment can disappear at any time, and it is not backed up.

What you type is processed

The demo is the real product. Text you enter is handled by that server, and where it reaches the assistant, the prompt goes to a model provider exactly as it would on your own install.

One rule

Do not enter real, personal or confidential information into the demo. It exists so you can look at the software without installing it, and nothing more than that.

The demo runs on the same machine as this site but under its own configuration, and it is deliberately excluded from the access log described above.

No telemetry, in any form

Mantle sends us no usage statistics, no crash reports, no feature analytics, no install ping and no licence check. We do not know how many installs exist, and we cannot find out.

Your data stays in your Postgres

Everything the brain knows — email, files, notes, pages, contacts, traces, embeddings — lives in the database on your machine. There is no synchronisation, no backup to us, and no key we hold.

Inference is the only traffic that leaves by default

Prompts and the context assembled for them go to the model providers you configure. Everything else on the wire is your own accounts, on your own credentials: your mailbox, your Microsoft tenant, your Telegram bot. The full list is on the architecture page.

Embeddings can stay on the machine

Local embedding via Ollama is supported and, when enabled, the text being embedded never leaves your hardware at all.

Updates are ordinary public downloads

Pulling a new release fetches public container images and public source. Those requests reach the registry and GitHub, under their logging and their policies, and they carry nothing about your brain.

You can verify all of this rather than believe it. The source is open, the deployment is a Compose file you can read, and every container is listed in it. If you want the wire-level version, the architecture documentation names every outbound connection the system can make.

If you point Mantle at a mailbox, a document store or a set of contacts, you are ingesting information about other people. For that data you are the controller, or the responsible party, and the obligations that come with it — a lawful basis, retention, honouring access and deletion requests, breach notification — are yours. We are not a processor for it, because we never see it.

Practically, that means the security of your install is also yours: your server, your TLS, your tokens, your model provider keys, your decisions about which agents get which tool grants and which peers get which shares. We cannot access your install, cannot recover anything from it, and cannot revoke anything on your behalf.

Self-hosting is what makes the privacy real. It also means nobody else is holding the other end of it.

Write to licensing@crossworks.engineering and we will respond within thirty days. There is no charge.

The honest limit: for this website there is almost certainly nothing of yours to return. The only data we hold is a masked log line with no identifier, no cookie and no account attached, and we have no way to determine which lines relate to you. Being unable to look you up is the intended design, not an evasion — but it does mean an access request will most often be answered with “we hold nothing that identifies you,” and we would rather say that here than surprise you with it later.

If you believe we have handled your information wrongly, you may complain to South Africa’s Information Regulator, or to the supervisory authority in your own country if you are in the EU or UK. We would prefer you tell us first, but that is your choice and not a precondition.

Children. This site is not directed at children and collects nothing from anyone, of any age, beyond what is described above.

International transfers. We do not transfer personal information to third parties. The masked log stays on the single server that produced it.

Cross Works Engineering (Pty) Ltd · licensing@crossworks.engineering

This page is dated at the top. If what we do changes, the page changes and the date changes with it. If a change is material — if we ever start collecting something we do not collect today — it will be described in plain language rather than absorbed into a paragraph, because a policy that quietly grows a new clause is how these documents lose their point.

The terms that go with this page are at mantle-ai.tech/terms.