Nothing on the proxy is authenticated, so whoever can reach the port can spend the user's PPQ.AI balance — true of every version, and not something either doc said. Enabling PPQ_DATA_DIR adds a second unauthenticated route, POST /setup/api-key, which replaces the stored key (upstream PayPerQ/ppq-private-mode-proxy#23). README states both and whose each is; instructions.md tells the user plainly to expose the interface only to people they trust, which is the part they can act on. TODO.md is deleted rather than emptied. Its live-inference item is closed — PayPerQ confirmed billed end-to-end against glm-5-2 and llama3-3-70b — and the other two are now issues #2 (backup/restore, unverified: it needs the server password, and the 0.4 file split gives it two new ways to fail) and #3 (bump when the auth patch lands, and retract the warnings it makes false). The scaffold's worklist has been worked; a file left behind to say so is not worth reading.
3.4 KiB
PPQ Private Mode
Documentation
- Upstream proxy documentation — endpoints, client examples, and how the encryption works
- PPQ.AI API docs — creating an API key and funding your account
What you get on StartOS
A local endpoint your AI clients talk to instead of talking to a model provider directly. Everything sent through it is encrypted on your server and only decrypted inside an attested hardware enclave, so neither PPQ.AI nor the model host can read your queries.
Two interfaces share one address:
- Status Page — a web page showing whether enclave attestation succeeded, which private models are available, and how to point a client at the proxy. It is also where you can save your API key.
- OpenAI-Compatible API — the same address, ready to copy into a client.
The only thing the service stores is your PPQ.AI API key and the logging setting.
Getting set up
You need a funded API key from ppq.ai — create one under Settings → API Keys, since usage is billed per query to that key. You can save it before or after starting the service; the proxy runs either way and simply refuses inference requests until it has one.
- Start the service. It performs a cryptographic attestation of the remote enclave, which the health check reports as ready — usually within a few seconds.
- Open the Status Page, confirm attestation says verified, and paste your key into the API key form. The proxy picks it up immediately, then restarts once to persist it.
- Copy the OpenAI-Compatible API address to point a client at it.
If you would rather not start the service first, run the Configure PPQ API Key action instead of step 2 — it does the same thing from the Actions tab.
Using PPQ Private Mode
Connecting a client
Point any OpenAI-compatible client at the API address:
POST <api-address>/v1/chat/completions
{"model": "private/glm-5-2", "messages": [{"role": "user", "content": "Hello"}]}
Anthropic-SDK clients (including Claude Code) can use the native endpoint at
<api-address>/v1/messages. For Claude Code, set:
export ANTHROPIC_BASE_URL="<api-address>"
export ANTHROPIC_MODEL="private/glm-5-2"
GET <api-address>/v1/models lists the available private models, and the Status
Page shows the same list.
Requests are billed to your saved API key by default. To bill a different
PPQ.AI key per request, pass it as an Authorization: Bearer header.
Replacing the key
Either surface works, and both write the same place: Replace key on the Status Page, or the Configure PPQ API Key action. The action is also where you turn Verbose Logging on and off; leaving its key field blank keeps the key you already saved.
Limitations
The proxy has no login of its own, so only expose it to people you trust. Anyone who can reach the address can send requests that are billed to your PPQ.AI key, and can replace the stored key from the Status Page. On a LAN-only address that means everyone on your network; if you enable a Tor or public address, it means anyone with the address. Keeping it on the LAN, or on Tor with the address kept private, is the safe default.
Backups of this service include your saved API key, and restoring a backup restores it — keep that in mind when handling backup drives.