PPQ Private Mode on StartOS
Everything not listed in this document should behave the same as upstream PPQ Private Mode. If a feature, setting, or behavior is not mentioned here, the upstream documentation is accurate and fully applicable — see the Documentation section of
instructions.mdfor links.
PPQ Private Mode is a local proxy that encrypts AI requests on your server and only lets them be decrypted inside an attested hardware enclave, so neither PPQ.AI nor the model host can read them. This package runs that proxy and points any OpenAI- or Anthropic-compatible client at it.
- Upstream proxy repo: https://github.com/PayPerQ/ppq-private-mode-proxy
- Upstream wrapper (this is a fork of it): https://github.com/Start9-Community/ppq-private-mode-startos
Changes in this fork
Forked from Start9-Community/ppq-private-mode-startos. The proxy behaviour is unchanged from upstream — the difference is how same-server clients such as Open WebUI connect to it:
- New action — Service-to-Service URL. Resolves the internal LXC-bridge address and prints the
http://<bridge-ip>:<port>/v1base URL that Open WebUI (or any service container on this same StartOS server) should dial — plain HTTP, no TLS or cert handling. This is what an on-box client must use instead of the LANhttps://address, which fails certificate validation inside another container's runtime. - Stable consumer contract.
startos/utils.tsnow exportsapiHostId('main') andapiPort(8787) for dependents to import, mirroring the pattern Open WebUI already uses for its other AI backends. - Docs.
README.mdandinstructions.mddocument the same-server / Open WebUI connection flow.
Everything else behaves like the upstream wrapper: version 0.4.1:1, built with SDK 2.0.9.
Table of Contents
- Image and Container Runtime
- Volume and Data Layout
- File Models
- Dependencies
- Network Access and Interfaces
- Installation and First-Run Flow
- Actions
- Tasks
- Health Checks
- Backups and Restore
- Limitations and Differences
- Quick Reference for AI Consumers
Image and Container Runtime
One image, built from upstream's own Dockerfile.
| Property | Value |
|---|---|
| Image | Built from the vendored upstream submodule |
| Architectures | x86_64, aarch64 |
| Command | The proxy's server entrypoint |
| Subcontainer | Purpose |
|---|---|
proxy-sub |
The only daemon — the one to attach to |
Upstream is vendored as a git submodule and built here, rather than pulled as a published image — so the shipped bits correspond to a specific upstream commit.
One oneshot runs first, giving the data directory to the unprivileged user the image runs as.
Volume and Data Layout
One volume, with a deliberate split down the middle.
| Volume | Subpath | Mount Point | Purpose |
|---|---|---|---|
main |
proxy |
/data |
Upstream's own data directory |
| Path | Written by | Holds |
|---|---|---|
proxy/config.json |
The proxy, and the action | The PPQ.AI API key |
store.json |
The action | The package's own settings |
Only the proxy subpath is mounted into the container. The package's own settings live at the volume root, outside what the proxy can see — so the proxy neither reads them nor overwrites them when it rewrites its own file.
File Models
Two models, and keeping them apart is the point.
| File | Format | Modelled | Written by |
|---|---|---|---|
proxy/config.json |
JSON | Yes — FileHelper.json |
The proxy, and the action |
store.json |
JSON | Yes — FileHelper.json |
The action |
The proxy's config file is co-owned, and its shape must stay exactly upstream's. The proxy rewrites the whole file when a key is saved on its status page, so anything else stored there would be destroyed — which is why the verbose-logging setting lives in a separate file the container cannot see.
The API key is not passed as an environment variable, though upstream supports that. Leaving it out means the proxy's own key store is the single owner of the key, so a key set through the status page and a key set through the action are the same key rather than two competing sources. The config file is read reactively, which is what makes a key saved by the action take effect: the write restarts the daemon, and the proxy reloads the file at start.
Dependencies
None.
But the service is not self-contained. Every request is forwarded to PPQ.AI's enclave, so the package needs working internet and a funded PPQ.AI account. Running this proxy locally is not running inference locally — what it buys is that the request contents are encrypted end-to-end into the enclave.
Network Access and Interfaces
Two interfaces on one address, which is unusual and deliberate.
| Interface | Id | Type | Port | Description |
|---|---|---|---|---|
| Status | ui |
ui | 8787 | Attestation state, available models, setup snippets |
| API | api |
api | 8787 | The OpenAI-compatible endpoint |
Both are exported from the same binding, because they are the same server: one is the page a person opens, the other is the URL a client is pointed at. Bound on the main MultiHost over HTTP and not masked.
Neither is authenticated by StartOS, and the status page is where the API key can be saved — so anyone who can reach the address can read the attestation state, use your key for inference, and replace it. Treat the address as a credential.
Same-server consumption (service-to-service). A client running in another service's container on this same server — Open WebUI, a custom agent — must not use the LAN https:// address: the calling runtime does not trust this server's self-signed Root CA, so the dial fails certificate verification. What it should use is the LXC-bridge address, which is plain HTTP — no TLS, no cert handling. Run the Service-to-Service URL action and paste the returned http://<bridge-ip>:<port>/v1 URL into the client. The LAN https:// address (e.g. https://192.168.0.158:56849) remains correct for off-box clients — a browser with the server's Root CA trusted, or curl -k.
Installation and First-Run Flow
Install seeds nothing and raises an important task asking for the PPQ.AI API key.
That task is deliberately not critical, and the reasoning is worth knowing: a critical task suspends the ordinary controls, which would leave the user unable to start the service — and the service's own status page is the other place a key can be saved. Blocking startup would block one of the two ways to complete the task.
So the proxy starts without a key. It serves its status page, performs attestation, and answers inference requests with an authentication error until a key exists.
A key saved on the status page clears the task too. The action clears its own task when it runs, but the status page never goes through the action — so the restart that the save triggers is where the package notices a key now exists and clears the prompt itself. Without that, StartOS would keep asking for a key the user had already set.
Actions
Two actions, both available whether or not the service is running.
Configure PPQ API Key
Sets the API key and the verbose-logging switch.
- What it changes: the key in the proxy's config file, and the logging setting in the package's own store.
- Cost: the service restarts, since the proxy reads its config at start.
- Repeat safety: idempotent. Leaving the key blank keeps the existing one rather than clearing it, so the action can be used to toggle logging without re-entering the key.
- The key is never echoed back into the form. The logging toggle is pre-filled; the key is not.
- The key's shape is checked before it is written. Upstream's status page rejects a malformed key outright, but a bad key written straight into the config file would only surface later as a failed request — so the same check is applied here.
Requests are billed to whichever key is set.
Service-to-Service URL
Resolves the internal LXC-bridge address another service on this same server (Open WebUI) dials to reach the proxy.
- What it returns: the copyable base URL
http://<bridge-ip>:<assigned-port>/v1. - Why it exists: the LAN
https://address fails from another container, because that container's runtime does not trust this server's Root CA; the bridge address is plain HTTP and needs no trust setup. - Usage: in Open WebUI, Admin Panel → Connections → OpenAI API, paste the URL, Save, and add the
private/model IDs shown on the Status Page. - Cost: none; the service does not restart.
Tasks
One, and it is advisory.
| Task | Severity | Raised when | Cleared when |
|---|---|---|---|
| Configure PPQ API Key | important |
An init that finds no key | The action runs, or a key appears |
important does not block the service from starting — see Installation and First-Run Flow for why that matters here.
Health Checks
One check, on the only daemon.
| Check | Displayed as | Method |
|---|---|---|
proxy |
"Encrypted Proxy" | Port 8787 is listening |
Here a port check means more than usual. The proxy binds its port only after remote attestation of the enclave has succeeded, so a listening port is evidence that the encrypted channel was verified — not merely that a process started.
It still says nothing about the key: an unset or rejected key shows a green check and an authentication error on the request.
Backups and Restore
The main volume is copied wholesale — sdk.Backups.ofVolumes('main'). That is the proxy's config file, holding the API key, and the package's settings.
The backup contains the API key in recoverable form, and requests made with it are billed to your account. There is nothing else here: no conversation history, no cache, no model data.
A restored instance comes back with the same key and works immediately, since nothing in the configuration is tied to the server it ran on.
Limitations and Differences
- Inference is not local. Requests go to PPQ.AI's enclave; the privacy claim is about who can decrypt them, not about where they run.
- A funded PPQ.AI account is required, and requests are billed to the configured key.
- Neither interface is authenticated, and the status page can both use and replace the key.
- The status page and the API share one address, so they cannot be exported separately.
- The API key cannot be cleared from the action — leaving the field blank preserves it.
- The package cannot add settings to the proxy's config file, which upstream rewrites wholesale.
- The only package-level setting is verbose logging. Everything else is upstream's.
- Same-server clients must use the bridge address, not the LAN
https://URL — see Network Access and Interfaces.
Quick Reference for AI Consumers
package_id: ppq-private-mode
image: built from ./ppq-private-mode-proxy # upstream vendored as a git submodule
architectures:
- x86_64
- aarch64
subcontainers:
- proxy-sub
volumes:
main:
proxy: /data # PPQ_DATA_DIR; store.json sits at the volume root, unmounted
file_models:
- proxy/config.json # upstream's shape exactly — the proxy rewrites it wholesale
- store.json # package-owned; kept outside the mount so the proxy can't clobber it
startos_managed_env_vars:
- PPQ_DATA_DIR
- HOST
- PORT
- DEBUG
# PPQ_API_KEY is deliberately NOT set — the proxy's key store owns the key
dependencies: [] # but requires internet and a funded PPQ.AI account
interfaces:
ui: { type: ui, port: 8787 } # status page; also where a key can be saved
api: { type: api, port: 8787 } # same binding, exported as a second interface
actions:
- configure-api-key # blank key field preserves the existing key
- show-service-to-service-url # prints http://<bridge-ip>:<port>/v1 for same-server clients (Open WebUI)
tasks:
- { action: configure-api-key, severity: important } # not critical, by design
health_checks:
- proxy # the port opens only after enclave attestation succeeds
Stable contract for dependents: startos/utils.ts exports apiHostId ('main') and apiPort (8787) — the values a package like Open WebUI imports rather than hardcoding (sdk.host.getBridgeAddress(effects, { packageId: 'ppq-private-mode', hostId: apiHostId, internalPort: apiPort, ssl: false })).