Astrixa

Astrixa is a containerized multi-provider LLM gateway with agent/provider registries, routing, guardrails, anonymization, authentication, and observability.

Components

• api-gateway: public entrypoint for chat/completions traffic
• gateway-proxy: proxy frontend for multi-replica gateway deployment
• routing-engine: provider selection and feedback handling
• provider-registry: provider metadata, pricing, limits, health state
• agent-registry: A2A agent records and auth metadata
• guardrails-engine: request and response policy checks
• anonymization-engine: local masking and de-anonymization
• auth-layer: bearer token validation and agent-scoped auth
• provider-adapters/mock-llm: local mock provider
• redis: shared admission and rate-limit state
• mlflow: run tracking
• prometheus, grafana, otel-collector, node-exporter: observability stack

Request Path

client -> gateway-proxy -> api-gateway -> auth-layer -> guardrails-engine
       -> anonymization-engine -> routing-engine -> provider
       -> response guardrails -> de-anonymization -> client

Capabilities

• multi-provider routing with mock and real providers
• streaming-safe proxying
• dynamic provider registry
• dynamic agent registry
• latency-aware and health-aware routing
• provider ejection and recovery
• ingress and response guardrails
• local anonymization with regex + spaCy NER
• independent policy_profile and anonymization_mode
• anonymization profiles: none, secrets-only, pii-lite, pii-strict, outreach
• anonymization controls: on / off, entity include/exclude, restore include/exclude
• agent-scoped policy inheritance
• admission control with global and per-provider concurrency limits
• Redis-backed shared admission and rate-limit state
• proxy-fronted multi-replica gateway deployment
• TTFT, TPOT, token, cost, auth, guardrail, and anonymization metrics
• Prometheus alerting and Grafana dashboards

Architecture

flowchart LR
    Client[Client / Agent] --> PX[Gateway Proxy]
    PX --> GW[API Gateway Replicas]
    GW --> AUTH[Auth Layer]
    AUTH --> GR[Guardrails Engine]
    GR --> ANON[Anonymization Engine]
    ANON --> RE[Routing Engine]
    RE --> PR[Provider Registry]
    RE --> AR[Agent Registry]
    RE --> PA[Providers]

    GW --> OTEL[OpenTelemetry]
    RE --> OTEL
    GR --> OTEL
    ANON --> OTEL
    AUTH --> OTEL
    GW --> REDIS[Redis]
    OTEL --> PROM[Prometheus]
    OTEL --> GRAF[Grafana]
    OTEL --> MLF[MLflow]

Loading

Docs

• deployment: docs/deployment.md
• product proposal: docs/product-proposal.md
• governance: docs/governance.md
• testing report: docs/testing-report.md
• routing comparison: docs/routing-strategy-comparison.md
• outage runbook: docs/runbook-provider-outage.md
• requirements matrix: docs/requirements-status.md
• submission checklist: docs/submission-checklist.md

Test Artifacts

• resilience checklist: tests/resilience/provider-ejection-checklist.md
• submission suite: tests/run_submission_suite.py