> ## Documentation Index
> Fetch the complete documentation index at: https://docs.acmeagentsupply.com/llms.txt
> Use this file to discover all available pages before exploring further.
# SphinxGate Overview
> Policy enforcement and auditable routing for multi-model AI agent systems
# SphinxGate
SphinxGate applies policy guardrails to your model routing. As agent systems grow — multiple models, fallback chains, different providers for different tasks — routing becomes a reliability surface in its own right. SphinxGate makes it controllable, inspectable, and provable.
## The Problem SphinxGate Solves
Multi-model agent systems have a routing problem that almost no one addresses until something goes wrong:
* **Which models are actually being called?** Not which ones you think — which ones are actually used in production.
* **Why did the fallback fire?** Was it expected or a silent failure?
* **Who approved this routing change?** Can you prove compliance?
* **Can you trust the audit trail?** If your routing logs are incomplete, you can't answer incidents or audits.
SphinxGate gives you policy-enforced, evidence-backed answers to all of these.
## What SphinxGate Does
Allow/deny rules for which models and providers can be used — and under what conditions.
Every routing decision logged with the model chosen, the reason, and the policy applied.
Configure fail-open vs. hard-fail behavior per route. No silent surprises.
Flags unexpected routing patterns — unauthorized models, unusual fallback frequency, policy violations.
## How Routing Policy Works
SphinxGate operates as a policy layer between your agent and the model router:
```
┌─────────────┐ ┌───────────────┐ ┌──────────────┐
│ Agent │────▶│ SphinxGate │────▶│ Model │
│ Request │ │ Policy Layer │ │ Provider │
└─────────────┘ └───────┬───────┘ └──────────────┘
│
┌───────▼───────┐
│ Routing Log │
│ + Evidence │
└───────────────┘
```
Every request passes through SphinxGate's policy evaluation before reaching a model. The decision — approved, denied, routed to fallback — is logged with full context.
## Policy Model
SphinxGate policies are simple allow/deny rules with conditions:
```yaml theme={null}
# ~/.acme/sphinxgate.yaml
policies:
- name: "production-models-only"
action: allow
models:
- anthropic/claude-*
- openai/gpt-4*
- name: "block-experimental"
action: deny
models:
- "*/preview"
- "*/experimental"
on_deny: hard_fail # fail the request; don't silently fallback
- name: "fallback-permitted"
action: allow
models:
- openai/gpt-3.5-turbo
conditions:
only_as_fallback: true # only permitted if primary denied
on_route: log_alert # log when this fires
```
### Fail Behaviors
| Behavior | Meaning | When to Use |
| ----------- | ------------------------------------- | ---------------------------------------- |
| `hard_fail` | Request fails with error | Compliance, security-sensitive routes |
| `fail_open` | Falls through to next policy/fallback | Resilience-first configurations |
| `log_deny` | Deny logged, request proceeds | Audit-without-blocking mode |
| `log_alert` | Permitted but flagged | Monitor unexpected-but-acceptable routes |
## Routing Log Format
Every decision is logged:
```json theme={null}
{
"timestamp": "2026-03-09T15:42:31Z",
"agent": "support-agent-prod",
"requested_model": "anthropic/claude-opus",
"routed_to": "anthropic/claude-opus",
"policy_applied": "production-models-only",
"action": "allow",
"fallback_fired": false,
"latency_ms": 4
}
```
This log is the evidence layer. Export it for audits, feed it to Agent911, or use it for your own observability.
## Anomaly Detection
SphinxGate flags routing patterns that don't match expectations:
| Anomaly | Trigger |
| ------------------------ | ------------------------------------------------ |
| Unauthorized model | A model outside the allow list was requested |
| Fallback frequency spike | Fallbacks firing much more than baseline |
| Policy violation attempt | Explicit deny rule triggered |
| Unknown provider | Request routed to unrecognized provider endpoint |
| Config bypass attempt | Routing attempted outside SphinxGate |
Anomalies are surfaced in the routing log and fed to Agent911 when configured.
## Quick Start
```bash theme={null}
# Check current routing policy
acme sphinxgate status
# View recent routing decisions
acme sphinxgate log --tail 20
# Validate a policy file
acme sphinxgate policy validate ~/.acme/sphinxgate.yaml
# Run in audit mode (log-only, no enforcement)
acme sphinxgate audit --agent my-agent-name
```
## CLI Reference
```bash theme={null}
# Status and overview
acme sphinxgate status
acme sphinxgate status --format json
# Policy management
acme sphinxgate policy show
acme sphinxgate policy validate
acme sphinxgate policy apply
# Routing log
acme sphinxgate log
acme sphinxgate log --tail 50 --follow
acme sphinxgate log --format json --output routing-log.json
# Anomaly report
acme sphinxgate anomalies
acme sphinxgate anomalies --since 24h
# Audit mode (enforcement off, logging on)
acme sphinxgate audit start
acme sphinxgate audit stop
```
## SphinxGate vs Agent-Level Config
You might already configure model fallbacks in your agent's gateway config. SphinxGate is **not** a replacement for that — it's a policy and audit layer on top of it.
| Layer | What It Does |
| -------------- | ------------------------------------------------------------ |
| Gateway config | Defines fallback chains and preferred models |
| **SphinxGate** | Enforces policy on top of those choices; logs every decision |
The distinction matters for compliance: your gateway config describes intent; SphinxGate provides evidence of what actually happened.
## Who SphinxGate Is For
SphinxGate is most valuable for teams that:
* Use **multiple models or providers** and need to know which ones fire
* Have **compliance requirements** around AI model usage
* Run **production-grade agent systems** where unauthorized routing is a risk
* Want **audit-ready logging** without building it themselves
## Pricing
SphinxGate is a paid product. See [Pricing](/pricing) for current rates.
## Next Steps
Signal transport integrity for the broader reliability stack.
SphinxGate anomalies feed into the Agent911 unified snapshot.
Runtime protection to complement routing governance.
Point-in-time reliability scan. Good first step.