ELUVIXSovereign AI Router

Native Tools & Agent Gateways

Eluvix is an agent-ready orchestration gateway. It allows developers and operators to bind external capabilities directly into model reasoning loops through native tools, Model Context Protocol (MCP) servers, and secure Agent Gateways.

Agent Gateway Node (agentGatewayNode)

The Agent Gateway acts as a security boundary between language models and external tool execution environments:

Access Policies

  • Restrictive (restrictive) (Recommended): Only explicitly connected and whitelisted tools can be called by models.
  • Permissive (permissive): Allows models to query approved internal VPC tool services.
  • Open (open): Connects to public MCP registries for broad capability discovery.

Tool Filtering Modes

  • Allow All (all): Exposes all connected tools to the LLM's function declaration manifest.
  • Strict Whitelist (whitelist): Filters declarations down to an approved subset of functions.
  • Blacklist Risky Tools (blacklist): Suppresses destructive tools (e.g. rm, drop_database, execute_arbitrary_shell).

Output PII Sanitization

When enabled, tool execution results are passed through the Privacy Guard before being returned to the model context. This prevents database queries or web scraper output from introducing external PII into subsequent reasoning turns.

Tool Execution Timeouts

Operators can configure granular execution timeout windows (e.g. 10,000ms) per gateway, preventing stalled third-party APIs from hanging agent loops.

OpenBao Secret Injection

Eluvix integrates with OpenBao / HashiCorp Vault secret engines (/api/secrets/discover). The Agent Gateway can automatically inject decrypted API keys or bearer tokens into outbound tool requests without exposing credentials in client canvases or client responses.

Model Context Protocol (MCP)

Eluvix natively supports the Model Context Protocol (MCP), the open standard for connecting AI systems to external data sources, enterprise tools, and local development environments.

MCP Tool Node (mcpToolNode)

The visual editor provides dedicated MCP Tool nodes:

  • Server Endpoint: Connects via Server-Sent Events (SSE) or local standard I/O (stdio).
  • Execution Modes:
    • Automatic (auto): Executes tool calls immediately upon model request.
    • Human-in-the-Loop (confirm): Halts execution and presents a confirmation modal before running potentially dangerous actions.
    • Dry Run (dry-run): Simulates tool execution and returns a mocked response for safe evaluation.
  • Parameter Validation: Automatically verifies model-supplied arguments against JSON schema definitions before network dispatch.

Native Built-in Tools

Eluvix includes zero-dependency native tools served directly from /api/v1/tools:

1. Calculator

  • Evaluates arithmetic and mathematical expressions securely without executing arbitrary code.
  • Ideal for small or fast reasoning models requiring exact numerical precision.

2. Web Scraper

  • Fetches external URLs, strips boilerplate HTML markup, and extracts clean, readable markdown or plain text.
  • Formats extracted content directly into the model's context for summarization or comparative analysis.