Skip to content

Model Routing Administration ​

This guide covers operating the Model Routing Protocol (MRP) from the Platform Admin dashboard.

Dashboard Overview ​

The MRP dashboard is accessible at Platform Admin > Model Routing > Live Routing. It has these tabs:

  • Status — MRP enabled/disabled, route counts, provider summary
  • Routes — Full MRIB route table with per-route attributes
  • Health Matrix — Per-provider health scores, circuit breaker states, maintenance controls
  • Events — Recent routing events (failovers, circuit breaks, health changes)
  • Shadow — Shadow-mode divergence report (when testing MRP before going live)
  • Controls — Route preferences, dampening, budget, and capacity settings
  • Topology — Visual network topology of providers and routes
  • Simulate — Single-model and bulk routing simulation, probes, and audits
  • Live — Real-time event stream and provider weathermap

Simulate Tab ​

Single Model Simulate ​

Enter a model ID (e.g., anthropic/claude-sonnet-4.6) and click "Run Simulation" to see how MRP would route a request. Model IDs are automatically normalized — vendor prefixes are stripped and aliases applied.

Simulate All ​

Click "Simulate All" to dry-run routing for every model registered in the MRIB. Results show which models resolve to exact matches vs family fallback vs no route. Green rows indicate exact matches, yellow indicates family-level fallback, red indicates no route found.

Probe All Models ​

Click "Probe All Models" to send a lightweight test prompt ("Say OK", max_tokens=5) to every reachable model. Results stream in real-time via SSE with a progress bar. Each result shows the model, provider, success/failure status, response latency, and a response preview.

Note: Probe All makes real API calls to every provider with configured keys. Each probe uses approximately 5 tokens. With many models configured, this may take 10-30 seconds.

Audit Context & Capabilities ​

Click "Audit Context & Capabilities" to compare hardcoded context window values against live provider data. The audit runs three layers:

  1. Static consistency — Checks that every model has both context window and max output token entries
  2. Live comparison — Compares hardcoded values against live provider API data
  3. Capability check — Flags routes with missing or drifted capabilities (vision, tools)

Severity levels: OK (green), Minor drift (yellow, <50% difference), Major drift (red, >2x difference), Missing (gray).

CLI Access ​

All diagnostic features are also available via the ghostai CLI:

bash
# Full suite (simulate + probe + audit)
ghostai mrp test-all

# Individual modes
ghostai mrp test-all --simulate-only
ghostai mrp test-all --probe-only
ghostai mrp test-all --audit

# Audit with severity filter
ghostai mrp test-all --audit --severity major

# JSON output for scripting
ghostai mrp test-all --json-output

Health Matrix ​

The Health Matrix tab shows per-provider health scores aggregated from the MRIB. Each provider card shows:

  • Health Score (0-1) — Weighted composite of latency, error rate, and availability
  • Circuit State — Closed (healthy), Half-Open (recovering), Open (tripped)
  • Latency P50/P99 — Median and tail latency in milliseconds
  • Error Rate — Fraction of recent requests that failed

Circuit Breakers ​

Circuit breakers automatically trip when a provider's error rate exceeds thresholds. In the Open state, no traffic is routed to that provider. The Half-Open state allows limited probe traffic to test recovery.

Maintenance Mode ​

Enter maintenance mode for a provider to gracefully drain traffic before planned downtime:

  1. Go to Health Matrix tab
  2. Click the maintenance toggle on a provider card
  3. Traffic will be rerouted to alternative providers

Exit maintenance mode to resume normal routing.

SOC 2 Type II Certified | BAA Available | HIPAA-Ready