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.
7.9 KiB
PPQ Private Mode on StartOS
Upstream docs: https://github.com/PayPerQ/ppq-private-mode-proxy
Everything not listed in this document should behave the same as the upstream PPQ Private Mode proxy. If a feature, setting, or behavior is not mentioned here, the upstream documentation is accurate and fully applicable.
End-to-end encrypted proxy for PPQ.AI private (TEE) AI models. Requests are encrypted on your server and only decrypted inside a hardware-secured enclave whose code fingerprint is cryptographically attested before any data is sent. Upstream: PayPerQ/ppq-private-mode-proxy.
Table of Contents
- Image and Container Runtime
- Volume and Data Layout
- Installation and First-Run Flow
- Configuration Management
- Network Access and Interfaces
- Actions (StartOS UI)
- Backups and Restore
- Health Checks
- Dependencies
- Limitations and Differences
- What Is Unchanged from Upstream
- Contributing
- Quick Reference for AI Consumers
Image and Container Runtime
The image is built from the upstream repository's own Dockerfile via the
ppq-private-mode-proxy git submodule (multi-stage node:22-alpine build,
runs as the unprivileged node user). No entrypoint modifications; the daemon
runs node /app/dist/bin/server.js directly.
Architectures: x86_64, aarch64.
Volume and Data Layout
One volume, main, split so the container sees only what it owns:
| Host path | Container path | Written by |
|---|---|---|
proxy/config.json |
/data/config.json |
the proxy (Status Page) and the SDK (action) |
store.json |
not mounted | the SDK only |
proxy/ is mounted at /data as the proxy's PPQ_DATA_DIR, and config.json
inside it is upstream's own file — the proxy rewrites it wholesale when a key is
saved from the Status Page, so nothing of ours may live there. The Verbose
Logging setting therefore lives in store.json, outside the mount.
A chown oneshot hands /data to the image's unprivileged node user before
the daemon starts, so the Status Page can write to it.
Installation and First-Run Flow
On install, an important (non-blocking) task prompts for a PPQ.AI API key. There are two ways to satisfy it, and neither is required before starting:
- Run the Configure PPQ API Key action.
- Start the service and save the key on the Status Page.
The proxy runs without a key — it completes attestation, serves the Status Page,
and answers 401 to inference requests until one is set. The task is
deliberately not critical: a critical task suspends every other control, which
would leave the user unable to start the service and reach the Status Page's own
key form.
Configuration Management
The API key is the proxy's to own. StartOS does not pass PPQ_API_KEY;
instead it points PPQ_DATA_DIR at the mounted volume so upstream's key store
is the single source of truth, written by either the action or the Status Page
and read from the same file by both.
PPQ_DATA_DIR—/data, enabling persistence and the Status Page's key formDEBUG— set from the action's Verbose Logging toggle, viastore.jsonHOST— fixed to0.0.0.0(container binding)PORT— fixed to 8787PPQ_API_BASE— not set (upstream defaulthttps://api.ppq.ai)
The proxy loads config.json once at startup, so main.ts watches the file:
writing a key restarts the daemon, which reloads it. A key saved on the Status
Page takes effect immediately in the running proxy and then triggers that same
restart — brief, and it converges on the value just saved either way.
Network Access and Interfaces
Two interfaces on one HTTP host, port 8787, both served by the proxy:
- Status Page (
ui) — upstream's server-rendered page at/: attestation state (enclave host and code fingerprint), API-key status, the private model list, and client setup snippets. - OpenAI-Compatible API (
api) — the same origin, shown as a copyable connection string for AI clients. Endpoints:GET /health,GET /v1/models,POST /v1/chat/completions(OpenAI format),POST /v1/messages(Anthropic format).
Availability on LAN/Tor is decided by the user from the interface panel.
No endpoint is authenticated. Upstream ships no auth on any route, so
whoever can reach the port can spend the user's PPQ.AI balance via
/v1/chat/completions, and — because this package sets PPQ_DATA_DIR — can
also replace the stored key via POST /setup/api-key
(upstream #23).
The billing exposure predates that endpoint and is inherent to the proxy; the
key-replacement one is enabled by our PPQ_DATA_DIR choice, and closes when
upstream's patch lands and this package bumps to it. instructions.md warns
the user to expose the interface accordingly.
Actions (StartOS UI)
- Configure PPQ API Key (
configure-api-key) — visible and enabled at any service status. Inputs: the PPQ.AI API key (masked, validated against upstream's^sk-[A-Za-z0-9]{16,64}$shape; left blank it keeps the key already saved and is never echoed back into the form) and the Verbose Logging toggle. No result output. Also the target of the first-run task.
Backups and Restore
The main volume (i.e. config.json) is backed up. Restore brings back the
saved API key and settings; no other state exists. Note that backups therefore
contain the saved PPQ.AI API key (StartOS backups are encrypted with the
user's backup password).
Health Checks
One daemon readiness check on port 8787. The upstream proxy only binds its port after remote attestation of the enclave succeeds, so "listening" implies the encrypted channel is verified. Until attestation completes the check reports the proxy as not ready.
Dependencies
None.
Limitations and Differences
- This service is a client-side encryption proxy, not a self-hosted model server: inference runs remotely in PPQ.AI's attested TEE enclaves, and a funded PPQ.AI API key is required (pay-per-query).
- Saving a key on the Status Page restarts the daemon a few seconds later.
The proxy applies the key immediately; the restart is StartOS reacting to
the same file write, and is what makes the Configure PPQ API Key action
work at all (upstream reads
config.jsononly at startup). - The upstream OpenClaw-plugin mode is not packaged; only the standalone proxy runs on StartOS.
- The bind address and port are fixed inside the container; upstream's
HOST/PORT/PPQ_API_BASEvariables are not user-configurable.
What Is Unchanged from Upstream
- All API endpoints, request/response formats, and streaming behavior
- The encryption and attestation flow (EHBP transport via the
tinfoilclient) - Model catalog and per-request
Authorization: Bearerkey override - The status page rendered at
/, its API-key form included
Contributing
See AGENTS.md.
Quick Reference for AI Consumers
package_id: ppq-private-mode
architectures: [x86_64, aarch64]
volumes:
main: /data # subpath `proxy/` only; `store.json` stays host-side
ports:
main: 8787 # one host serving both interfaces
interfaces:
- ui
- api
dependencies: none
startos_managed_env_vars:
- PPQ_DATA_DIR
- HOST
- PORT
- DEBUG
actions:
- configure-api-key