llm-analysis-assistant

Implements a streamlined Model Context Protocol (MCP) client that abstracts three distinct transport mechanisms: stdio (local process communication), SSE (Server-Sent Events for streaming), and streamable HTTP (bidirectional HTTP streaming). The client handles protocol negotiation, message serialization/deserialization, and transport-specific connection lifecycle management, allowing unified MCP interactions across heterogeneous server implementations without transport-specific client code.

Provides a web-based /logs page that captures and displays all MCP client requests and server responses in real-time, including request payloads, response bodies, latency metrics, and error details. The dashboard stores request history in-memory or persistent storage, enabling developers to inspect protocol-level interactions, debug integration issues, and audit MCP communication patterns without instrumenting client code.

Implements a mock OpenAI-compatible API endpoint that intercepts and logs requests matching OpenAI's chat completion and embedding API schemas, allowing developers to test client code against a local endpoint without consuming API credits. The simulator validates request format, tracks API usage patterns, and can replay recorded responses, enabling integration testing and behavior monitoring of OpenAI-dependent code.

Provides a mock Ollama API endpoint compatible with Ollama's chat and embedding endpoints, allowing developers to test Ollama-dependent code locally with configurable model responses. The simulator validates request format against Ollama API specifications, logs all interactions, and supports response templating for deterministic testing of LLM workflows without requiring a running Ollama instance.

Captures all MCP protocol messages across stdio, SSE, and HTTP transports into a unified request/response log, enabling developers to replay recorded interactions, analyze communication patterns, and test client behavior against deterministic server responses. The capture mechanism operates transparently at the transport layer, preserving timing information and streaming semantics without modifying client or server code.

Implements transport-specific streaming response handling for SSE and HTTP streaming transports, buffering partial messages, managing backpressure, and reassembling chunked responses into complete MCP protocol messages. The implementation handles transport-specific framing (SSE event boundaries, HTTP chunk encoding) while presenting a unified streaming interface to client code, abstracting away transport-level complexity.

Implements MCP-specific error handling that distinguishes between transport errors (connection failures, timeouts), protocol errors (invalid JSON-RPC format, missing required fields), and application errors (MCP server returning error responses). The system provides structured error context including error codes, messages, and recovery suggestions, enabling client code to implement intelligent retry logic and graceful degradation strategies.

Collects and aggregates metrics on all MCP requests including latency (p50, p95, p99), throughput, error rates, and per-endpoint statistics. Metrics are exposed through the /logs dashboard and can be exported for external monitoring systems. The collection mechanism operates transparently at the transport layer, capturing timing information without requiring client instrumentation.

Manages connection establishment, authentication, keep-alive, and graceful shutdown across stdio, SSE, and HTTP streaming transports. For stdio, handles process spawning and lifecycle; for SSE/HTTP, manages connection pooling and reconnection logic. The system provides unified connection state tracking and automatic reconnection with exponential backoff for transient failures.

Validates outgoing MCP requests against JSON-RPC 2.0 schema and transport-specific requirements before transmission, ensuring protocol compliance and catching client-side errors early. The system serializes requests to transport-appropriate formats (JSON for HTTP/SSE, newline-delimited JSON for stdio) and deserializes responses, handling type coercion and format conversion transparently.