Eluvix Sovereign AI Router & Prompt/PII DLP Guard
The eluvix plugin (sudo axonwallctl plugin enable eluvix) embeds the Eluvix Sovereign AI Router directly into your Axonwall perimeter (control plane on port 3588 and Switchyard OpenAI/Anthropic-compatible proxy on port 8080). Instead of allowing developer workstations, autonomous coding agents, and internal services to send raw source code, API keys, and PII directly to external cloud LLM APIs, Eluvix intercepts, sanitizes, caches, and routes every AI request through a zero-trust sovereign gateway.
1. Switchyard 4-Tier Complexity Routing
Up to 75% of enterprise LLM traffic consists of low-complexity boilerplate, JSON formatting, log summarization, and tool-call argument extraction. Eluvix's Switchyard classifier evaluates prompt complexity in sub-millisecond time and routes requests across four configurable tiers:
Routine classification, tool-call parsing, and code completion are routed to on-premise Ollama, vLLM, or Lemonade NPU/GPU inference nodes on your LAN—keeping sensitive data 100% on-premise at zero marginal token cost.
Deep architectural reasoning and multi-file synthesis are routed to frontier providers (Google Gemini, Anthropic Claude, OpenAI) only after passing through inline PromptGuard and PrivacyGuard DLP redaction.
2. Inline PromptGuard & Reversible PrivacyGuard DLP
Every inbound prompt and outbound tool observation passes through two deterministic security stages before reaching any external model:
Automatically detects and replaces AWS keys (AKIA*), GitHub tokens (ghp_*), OpenAI/Anthropic keys (sk-*), PEM private keys, internal RFC 1918 IPv4/IPv6 addresses, emails, and SSNs with deterministic vault placeholders (e.g. <ELUVIX_SECRET_01>) before cloud egress, and re-hydrates them in the streaming response.
Inspects multi-turn agent context and MCP tool stdout/stderr for indirect prompt injections, system-prompt extraction attempts, and covert markdown/URL data exfiltration channels.
3. Native Agent Harness Multiplexer (8 Harnesses)
Eluvix includes first-class protocol adapters and deep-fork runtime harnesses for autonomous coding and SecOps agents:
| Harness ID | Protocol / Port | Sovereign Integration Capabilities |
|---|---|---|
| opencode | HTTP SSE (:4096) | Inline DLP on tool stdout/stderr, KV-cache prefix pinning, zero-copy Axonwall MCP binding. |
| acp | JSON-RPC 2.0 stdio | Universal multiplexer for Antigravity (agy), Claude Code ACP, and Goose inside gVisor/Firecracker. |
| trueforge | In-Process ReAct (:3588) | Sub-ms Switchyard tier switching per reasoning step + Graphiti episodic memory graph. |
| openhands | Action Stream (:3000) | Sandbox egress enforced via Axonwall nftables + credential scrubbing on bash observations. |
| mastra / letta / a2a / hermes | :4111 / :8283 / :41241 / :4280 | Structured workflows, stateful memory blocks, Google A2A JSON-RPC peer cards, and multi-channel messaging. |
4. Declarative Configuration (eluvix: in /etc/axonwall/system.yaml)
Enable the built-in eluvix plugin via sudo axonwallctl plugin enable eluvix, then configure eluvix: inside /etc/axonwall/system.yaml:
eluvix:
enabled: true
deploymentMode: systemd # 'systemd' | 'docker' | 'k3s-helm'
listenPort: 3588
switchyardPort: 8080
domain: eluvix.axonwall.local
teePolicy: prefer-sev-snp
ldapAuth:
enabled: true
bindAccountCn: svc-eluvix
adminGroup: eluvix-admin
operatorGroup: eluvix-operator
readonlyGroup: eluvix-readonly
tokenomics:
strategy: hybrid-sovereign
semanticCacheEnabled: true
kvArchitecture: deepseek-mla-fp8
rdmaConnector: vllm-nixl
promptGuard:
enabled: true
blockIndirectInjection: true
inspectMcpToolStdout: true
privacyGuard:
enabled: true
reversibleTokenization: true
redactSecrets: true
redactPii: true
redactInternalIps: true
harnessMesh:
enabled: true
sandboxBackend: gvisor
memoryEngine: graphiti
injectAxonwallMcp: true
enabledHarnesses:
- opencode
- trueforge
- acp
- openhands
- mastra
- letta
- a2a
- hermes/ai/eluvix, in the Neovim Super Configurator via axonwall-config system, or verify status via axonwallctl status.