Platform & deployment

TrueFoundry + AWS Bedrock

The AI Gateway is the control plane for everything Halo does — every model call, every tool call, every guardrail.

Model Gateway

Every LLM call routes through the gateway to AWS Bedrock. Nothing talks to a model directly. This is what makes the rest of the resilience story possible — routing, fallback, and observability are all enforced at one layer.

Virtual Models

Halo defines two virtual models, priority-routed to Bedrock targets:

Virtual modelRole
halo-vm-normalPrimary Claude on Bedrock with a fallback target behind it.
halo-vm-degradedA faster / cheaper model used when Halo steps down a mode.

On a rate-limit, timeout, or 5xx, the gateway fails over to the next target automatically and the backend steps the agent down a mode. See Resilience & modes.

MCP Gateway

The same gateway governs the product's tools, exposed as two virtual MCP servers — jaguar-observe (read-only) and jaguar-act (approval-gated writes). Splitting the tools across two servers gives per-server governance out of the box.

Guardrails

A secrets-detection guardrail runs inline at the gateway, blocking sensitive input before it reaches a model — with no detection logic in our code. See Guardrails & safety.

Observability

Every step leaves a trace. Halo surfaces it right in the war room: which model resolved, how many spans the run produced, which guardrails fired, and exactly which tools the agent called. That is the audit trail that lets us prove a run happened rather than just claim it.

Heads-up from building on it: the spans query API was occasionally slow, so for the live demo Halo serves a cached trace summary instead of issuing a live span query mid-incident — keeping the demo deterministic.