MCP CLI Client

Manages the complete lifecycle of MCP server processes including startup, shutdown, and graceful termination. The CLI host spawns and monitors external MCP server processes, handling stdio-based bidirectional communication channels and ensuring proper resource cleanup. Implements process supervision with error handling for server crashes and connection failures.

Routes JSON-RPC 2.0 messages between the LLM client and MCP servers, handling request/response correlation, error mapping, and protocol-level concerns. Implements message framing over stdio with proper serialization/deserialization, timeout handling, and error response generation. Translates between LLM tool-calling conventions and MCP's standardized JSON-RPC interface.

Discovers available tools from connected MCP servers by querying their tool list endpoints and extracting JSON schemas describing tool parameters, return types, and documentation. Builds a unified tool registry that aggregates capabilities across multiple MCP servers, enabling the LLM to understand what tools are available and how to invoke them. Handles schema validation and normalization across different server implementations.

Provides a unified interface for invoking tools regardless of which LLM is making the request, abstracting away differences between OpenAI function calling, Anthropic tool use, Claude messages, and other LLM-specific conventions. Translates tool invocation requests from any LLM format into MCP JSON-RPC calls and maps responses back to the LLM's expected format. Handles parameter binding, type coercion, and result formatting.

Parses command-line arguments and binds them to MCP tool parameters, enabling direct invocation of tools from the shell. Implements argument parsing with support for flags, positional arguments, and complex data types (JSON objects, arrays). Maps CLI arguments to tool parameter schemas and validates types before invoking the tool through MCP.

Provides an interactive read-eval-print loop (REPL) for discovering, testing, and invoking MCP tools without writing code. Displays available tools with their descriptions and parameters, accepts tool invocation commands with argument completion, and formats results for human readability. Maintains session state and command history for iterative tool exploration.

Formats tool execution results into human-readable and machine-parseable output formats including JSON, YAML, table, and plain text. Implements custom formatters for different result types and supports filtering/projection of result fields. Handles large result sets with pagination and truncation to prevent terminal overflow.

Manages configuration for MCP server connections, CLI behavior, and tool invocation defaults through configuration files (JSON, YAML, TOML) and environment variables. Supports server definitions with connection parameters, authentication credentials, and tool filtering rules. Implements configuration inheritance and override precedence (CLI args > env vars > config file > defaults).

Captures and reports errors from tool invocations with detailed diagnostic information including stack traces, request/response payloads, and timing data. Implements structured logging with configurable verbosity levels and output destinations. Provides error recovery strategies such as automatic retries with exponential backoff for transient failures.

Caches tool execution results based on input parameters, avoiding redundant invocations of expensive or idempotent tools. Implements cache key generation from tool parameters, configurable TTL (time-to-live) for cache entries, and cache invalidation strategies. Supports both in-memory and persistent caching backends.