AiPipe Integration

AiPipe is an on-device LLM router that automatically selects the most cost-effective model for each request, reducing AI spend without changing how agents work.

What it does

• Smart routing: scores each request's complexity (length, code presence, keyword category, conversation depth) and picks the cheapest model that can handle it
• Multi-provider: routes across OpenAI, Anthropic, xAI (Grok), OpenRouter, MiniMax, Kimi, and Gemini based on cost and capability
• TTFT-aware streaming: for streaming calls, models with high time-to-first-token are penalised in selection
• Context window guard: excludes models whose max context window is smaller than the request's total token count
• Caching: identical requests return cached responses instantly
• Stats dashboard: live view of requests routed, cost saved vs GPT-4o, cache hit rate, provider breakdown (per-tenant isolated)

Dashboard

The ⚡ Router tab on the dashboard shows:

• Total requests routed
• Estimated cost savings vs always using GPT-4o
• Cache hit percentage
• Queue depth and p50 latency
• Provider breakdown bar chart
• Top model used

The widget refreshes every 30 seconds automatically.

Setup Wizard

Step 4 of the connection wizard ("Enable Smart Routing") checks whether AiPipe is running and explains how it uses your existing API keys from Step 3. The step is always skippable. AiPipe is optional.

Using the proxy

Your agents can route through AiPipe via the ArchonHQ API:

POST /api/aipipe/proxy/chat      # OpenAI-compatible
POST /api/aipipe/proxy/messages  # Anthropic-compatible

All proxy routes require authentication (Bearer token or session cookie). Invalid request bodies return 400 with a Zod validation message.

Health check

GET /api/aipipe/health

Returns {"status":"ok"} when AiPipe is reachable, or {"status":"unavailable"} with HTTP 503 when it is not.

Availability

AiPipe is optional. If the service is not running, the health endpoint returns 503 and the dashboard widget shows an offline state with a retry button. All other dashboard features continue to work normally.