Before
Resolve the Project and enforce route, model, CIDR, field, rate, concurrency, and budget policy.
PRODUCT
Turn model access into a governed local boundary.
Applications use a Gateway Key and public model alias. Halro owns the real credentials, target selection, budget reservation, retry evidence, and final usage settlement.
REQUEST LIFECYCLE
Controls before the call; evidence after it.
During
Select an eligible Deployment and preserve every Provider Attempt separately.
After
Settle actual tokens and cost with pricing evidence, latency, errors, and terminal state.
RESOURCE MODEL
Six resources connect a provider target to a business owner.
- Credential
Stores provider credentials away from application code.
- Provider
Defines the access surface, protocol adapter, and capability evidence.
- Deployment
Pins the actual model target, version, and pricing evidence.
- Route
Exposes a stable alias and binds allowed Deployments.
- Project
Owns budget, rate, concurrency, security, and cost attribution.
- Gateway Key
Gives each application an independent identity and revocation boundary.
CURRENT LIMITS
Evaluate the operating model as it exists today.
Deployment: standalone single writer; this is not presented as a highly available cluster.
Scope: Halro is not an agent orchestrator, workflow engine, memory system, or evaluation platform.
Protocol: OpenAI- and Anthropic-facing APIs are documented per endpoint. Realtime WebSocket/WebRTC is not claimed.