Secure access to Drabsnarg AI

Objective

Make the existing local AI application available to explicitly approved users through ai.drabsnarg.com, while retaining individual accounts and private conversation histories.

Project status

This is a planned infrastructure project. The architecture is defined; this public write-up does not confirm that the tunnel, Access policy, or invited-user setup is operational.

Requirements

Reuse the working native Open WebUI and llama.cpp services.
Require secure authentication before reaching the application.
Disable uncontrolled account registration.
Keep application administration restricted.
Preserve separate accounts and conversation histories.
Avoid opening inbound router ports.
Keep inference endpoints inaccessible from the public internet.

Proposed architecture

Approved user
    ↓  HTTPS
Cloudflare Access
    ↓
Cloudflare Tunnel
    ↓
Open WebUI
    ↓  Local-only inference connection
llama.cpp → Qwen

Cloudflare Access controls who can reach the application. Open WebUI controls application accounts and permissions. These are separate boundaries and need separate validation.

Implementation approach

First inventory installed versions and back up the existing application. Configure the tunnel to expose only the intended application, then apply an explicit allow policy in Cloudflare Access. Verify application registration and role settings before inviting additional users.

Any use of identity headers must follow the installed version’s documented trust model. An identity header supplied by an arbitrary client must not grant application access.

Acceptance checks

Scenario	Required behavior
Unapproved visitor	Cannot reach the chat application
Approved user	Can authenticate and use their own account
Two invited accounts	Cannot access each other’s private histories
Standard user	Cannot reach administrative functions
Direct inference request	No public route to the backend
Access removal	Re-entry and session behavior match the removal policy
Android and desktop browsers	Sign-in and chat remain usable
iMac offline	Public portfolio still works

Recovery and limitations

Document application backups, restoration, tunnel startup, and restart behavior. The iMac is a limited-capacity personal server; access does not imply guaranteed availability or fast responses.

Future improvements

After authentication and isolation are verified, evaluate concurrency limits and basic operational monitoring. Programmatic API access remains a separate project.