Cost And Provider Strategy
Prox OS should be BYOK first and hosted credits later. It should not promise
Principle
Prox OS should be BYOK first and hosted credits later. It should not promise unlimited free AI.
The product should reduce model cost by turning repeated work into structured workflows, local actions, cached artifacts, and explicit restore policies.
BYOK First
Free and early prototype plans can allow users to bring their own model key. The current Alma settings dialog is a local development prototype only.
Production BYOK requires:
- secure vault
- secret redaction
- provider adapter boundary
- no secret exposure to model context
- audit of key usage metadata without storing raw key material in logs
Hosted Credits Later
Hosted inference should use AI Credits or AI Action Units.
AI Action Units can represent:
- model calls
- workflow runs
- artifact generation
- workspace generation
- agent review steps
- hosted tool execution
This avoids vague unlimited AI claims and gives heavy users a path to BYOK or top-up.
Model Routing
Future model router should consider:
- small model for classification, summaries, routing, and UI copy
- larger model for workspace generation, app spec generation, and complex reasoning
- local model for private or low-cost tasks
- provider availability
- user policy
- cost ceiling
- latency target
Workflow-First Cost Reduction
Workflow Cards should run local OS actions whenever possible:
- open apps locally
- restore local layout locally
- render deterministic artifact previews locally
- ask for confirmation locally
- call a model only when reasoning or generation is needed
Non-goals In Current Task
- No SDK integration.
- No provider adapter.
- No model router.
- No hosted credit ledger.
- No billing route.
- No API key vault.