@mcptoolgate/client

Intercepts MCP tool invocations from Claude Desktop before execution and routes them through a human approval workflow. Implements a middleware pattern that sits between the MCP client and tool handlers, capturing tool calls, presenting them to a human reviewer with full context (tool name, parameters, description), and only allowing execution upon explicit approval. Uses event-driven architecture to maintain non-blocking async approval flows.

Captures all outbound MCP tool calls from Claude Desktop at the protocol level and enriches them with metadata before routing to approval or execution. Implements a transparent proxy pattern that parses MCP messages, extracts tool invocation details (name, parameters, schema), and augments them with execution context (timestamp, caller identity, risk classification). Maintains full fidelity of original tool definitions and parameter types for accurate approval decisions.

Maintains a persistent record of all tool approval decisions, rejections, and execution outcomes with full audit trail metadata. Implements append-only logging with immutable records including approver identity, decision timestamp, tool details, parameters, and execution result. Supports structured query and export of approval history for compliance reporting and forensic analysis. Uses event sourcing pattern to ensure audit trail integrity.

Manages the lifecycle of MCP server connections from Claude Desktop, including connection establishment, health monitoring, graceful shutdown, and error recovery. Implements connection pooling with automatic reconnection logic and heartbeat monitoring to detect stale connections. Handles MCP protocol handshake, capability negotiation, and tool definition discovery. Provides hooks for custom connection policies and rate limiting per MCP server.

Provides a user interface for reviewing and approving/rejecting tool invocations, integrated with Claude Desktop's native UI or presented via a companion web interface. Displays tool name, description, parameters with their values, and risk classification. Implements approval decision capture with optional comments and reason codes. Uses real-time notification to alert users of pending approvals and push decisions back to Claude Desktop execution context.

Automatically classifies MCP tools by risk level (low, medium, high, critical) based on tool metadata, parameter types, and configurable risk policies. Implements rule engine that applies different approval workflows based on risk classification — low-risk tools may auto-approve, medium-risk require single approval, high-risk require multi-level approval. Supports custom risk scoring functions and policy definitions in declarative format. Enables dynamic rule updates without restarting the client.

Enables multiple users to participate in approval workflows with role-based access control (RBAC) and approval authority delegation. Implements role definitions (approver, reviewer, auditor) with granular permissions (approve high-risk tools, view audit logs, modify policies). Supports approval routing rules that assign pending approvals to specific users or groups based on tool category or risk level. Tracks approval authority and enforces approval quorum for critical operations.

Validates all tool invocation parameters against their declared JSON Schema definitions before approval or execution. Implements schema validation with detailed error reporting for type mismatches, missing required fields, and constraint violations. Supports custom validation rules and parameter sanitization logic. Prevents execution of tool calls with invalid parameters, protecting downstream systems from malformed requests.