The current setup uses the Model Context Protocol (MCP) which is:

• Request-Response based: Each tool call is a single request that returns a complete response
• JSON-based messaging: Structured but not optimized for streaming
• Stdio transport: Communication over stdin/stdout pipes
• Synchronous by nature: Even with async Python, the protocol itself expects complete responses

Worth noting the system prompt is where lot of editing/polishing is needed:

• parse user intent system prompt is key
• {tools_context}
• decision logic
• tools selection rules
• parameter validation and confirmation
• file input handling
• portfolio file handling pa tools
• spar parameter handling
• univese parameter handling
• realtioship parameter handling
• save/export handling
• response behavior response format
• EXAMPLES

If the user asks a GENERAL KNOWLEDGE question (about companies, facts, definitions, etc.), respond with: {{"type": "general_answer", "answer": "your answer here"}}
remember previous conversations and use context

Let’s take a deeper look at the fundamental infrastructure, if we need to apply better communication experience for users.

First, I have intelligent_client.py and intelligent_client_tcp.py, let’s compare

spect	intelligent_client.py	intelligent_client_tcp.py
MCP Transport	STDIO pipes (subprocess)	Raw TCP sockets
Connection	Local process spawning	Network connection
Protocol	MCP over STDIO	JSON-RPC over TCP
Deployment	Single machine	Distributed (over VPN)

The key is uses STDIO pipes is for local MCP communication and HTTPS for Azure OpenAI, while the TCP version uses raw TCP sockets for remote MCP communication. At present, for a small team, STDIO is sufficient.

Further, I’d like to explore gRPC over JSON-RPC,

Aspect	JSON-RPC (Current)	gRPC
Message Format	Text-based JSON	Binary Protocol Buffers
Schema Definition	Loose/Dynamic	Strict .proto files
Type Safety	Runtime validation	Compile-time validation
Performance	Slower (JSON parsing)	Faster (binary serialization)
Human Readability	Easy to debug/inspect	Requires tools to inspect
Streaming	Request/Response only	4 streaming patterns
Code Generation	Manual implementation	Auto-generated clients

gRPC is more for performance optimization, hence in my case, json-RPC is a good solution.