NET · Gemini workspace
How the Mercatus Blueprint .NET robot talks to Gemini — a four-stop guided tour with sample request/response JSON, plus captures from the Google AI Studio multimodal workspace session. Want the real thing? Chat with Assist (LIVE when the VPS is healthy) →
Not the live .NET API. This page is a static explainer: the JSON below is sample data, not a live call. The Mercatus Blueprint Robot itself cannot run on the DreamHost shared host (CoreCLR SIGSEGV / runtime create failure), so it runs on the Mens Mercatus VPS and Assist reaches it through a same-origin proxy. The AI Studio app in the screenshots is the multimodal workspace (balance sheet, operational logs, compliance badges, NetzarLab, governance, NZC→NZG token rebalancing); figures in the guide are demo state inside that app.
Two pieces: a static portfolio on DreamHost shared hosting, and an ASP.NET Core 8 minimal API (MercatusCartelRobot) on the VPS. The robot exposes three endpoints and loads its system prompt from MERCATUS_CARTEL_SYSTEM_PROMPT.md at boot.
The AI Studio workspace (screenshots below) is a separate multimodal prototype — same Blueprint brief, different runtime.
The robot uses an OpenAI-compatible chat/completions body (model + temperature + messages) so the provider is swappable via config (Llm:BaseUrl, Llm:Model, Llm:ApiKey — the key lives in server env, never in this page). Retries: up to 4 attempts on 429/502/503/504 or "UNAVAILABLE / high demand" bodies, exponential backoff ~0.6 s → 1.2 s → 2.4 s with jitter.
Pick an endpoint and scenario to see the request and response shapes:
The HTTPS portfolio never calls a raw HTTP IP. Mercatus Assist probes health, then routes chat through a same-origin PHP proxy:
If /health fails, Assist mounts DEMO keyword coaching instead — the chat never dead-ends. Open Assist →
llmConfigured as a boolean, never the key.Captions only. No narration. Same three frames.