mcp-runtime-guard

Intercepts MCP tool invocations at runtime and validates them against declarative policy rules before execution. Implements a proxy pattern that sits between the MCP client and server, parsing tool call requests, matching them against policy conditions (tool name, arguments, caller identity), and either allowing, denying, or modifying the call based on policy evaluation. Uses a rule-matching engine to enforce fine-grained access control without modifying underlying tool implementations.

Validates MCP tool call arguments against schema constraints and optionally transforms or sanitizes arguments before tool execution. Likely uses JSON Schema or similar validation to check argument types, ranges, and formats, with support for custom validation rules defined in policy. May include argument filtering (removing sensitive fields) or normalization (converting formats) based on policy directives.

Evaluates tool call permissions based on caller identity (user, model, application) and request context (source IP, timestamp, session). Implements identity-aware policy evaluation where rules can reference caller attributes and context metadata to make access decisions. Likely uses a context object passed through the MCP request to identify the caller and evaluate policies conditionally based on identity attributes.

Provides a declarative policy language or configuration format for defining tool access rules, validation constraints, and transformation logic. Likely uses a structured format (YAML, JSON, or custom DSL) to express policies as rules with conditions and actions. Includes mechanisms for loading, parsing, and evaluating policies at runtime, with support for rule composition and precedence.

Logs all tool invocations (allowed, denied, modified) with metadata including caller identity, tool name, arguments, decision reason, and timestamp. Implements structured logging that captures the full context of each tool call decision, enabling audit trails and monitoring. Likely writes logs to stdout, files, or external logging services in a structured format (JSON or similar).

Rejects tool calls that violate policy rules and returns standardized error responses to the caller. Implements a denial mechanism that prevents tool execution and communicates the denial reason (policy violation, validation failure, access denied) back through the MCP protocol. Likely returns MCP error responses with structured error details and policy violation reasons.

Routes MCP requests through the proxy, parsing JSON-RPC messages, extracting tool call information, and forwarding validated requests to the underlying MCP server. Implements a transparent proxy that intercepts MCP protocol messages, applies policy evaluation, and forwards requests while maintaining protocol semantics. Handles both request and response routing, ensuring that tool responses are returned to the caller correctly.