LLM-Persistent-Agent-System-Service (MCP Implementation)

This is a TypeScript-engineered service framework conforming to the Model Context Protocol (MCP) specification, designed to endow LLMs with robust, externalized memory capabilities. It exposes utilities for interactive dialogue sessions and seamless orchestration across disparate LLM processing units.

Core Functionalities (Tools)

• converse - Transmit an input utterance to the currently designated LLM backend.

  • Requires a mandatory utterance payload.
  • Supports service endpoints including OpenAI implementations, Anthropic models, OpenRouter gateways, and local Ollama instances.

• fetch_session_log - Retrieve chronologically stored interaction records.

  • Accepts an optional count_cap argument to cap the number of records returned.
  • Setting count_cap: null forces retrieval of the entire preserved history.
  • Results are structured chronologically, appended with time markers.

• purge_session_log - Completely erase all previously recorded conversational artifacts.

  • Deletes the entirety of the session's persistent data.

• delegate_backend - Dynamically reassign the operational LLM engine.

  • Valid targets include OpenAI, Anthropic, OpenRouter, and Ollama.
  • The chosen backend selection is persistently stored across sessions.

• reconfigure_model - Modify the specific computational model utilized by the active backend engine.

  • Model specifications are contingent upon the selected provider:
    • Anthropic Offerings:
      • Claude 3 Lineup:
        • claude-3-haiku: Optimized for maximal throughput and minimal latency in high-volume scenarios (e.g., support queues).
        • claude-3-sonnet: Engineered for optimal general-purpose task execution.
        • claude-3-opus: Tailored for complex inferential logic and peak performance requirements.
      • Claude 3.5 Lineup:
        • claude-3.5-haiku: Improved efficiency and cost-performance profile.
        • claude-3.5-sonnet: Superior analytical capabilities, especially in code-centric or interactive simulation tasks.
    • OpenAI Parameters: Supported identifiers such as 'gpt-4o', 'gpt-4o-mini', 'gpt-4-turbo'.
    • OpenRouter Formats: Accepts arbitrary 'vendor/model' strings (e.g., 'mistralai/mixtral-8x7b').
    • Ollama Targets: Accepts any model name locally accessible by the Ollama runtime (e.g., 'solar', 'llama3:8b').
  • The selected model configuration is maintained throughout the service lifecycle.

Deployment & Setup

Dependency installation prerequisites: bash npm install

Compiling the service binary: bash npm run build

For continuous development integration (automatic recompilation): bash npm run watch

Integration Instructions (For Claude Desktop)

To link this service, integrate the following configuration snippet into your application's local settings file:

MacOS Path: ~/Library/Application Support/Claude/claude_desktop_config.json Windows Path: %APPDATA%/Claude/claude_desktop_config.json

Configuration Structure:

{ "mcpServers": { "persistent-llm-hub": { "command": "/path/to/LLM-Persistent-Agent-System-Service/build/index.js", "env": { "OPENAI_API_KEY": "Your_Secret_Key_For_OpenAI", "ANTHROPIC_API_KEY": "Your_Secret_Key_For_Anthropic", "OPENROUTER_API_KEY": "Your_Secret_Key_For_OpenRouter" } } } }

Required Runtime Secrets (Environment Variables)

• OPENAI_API_KEY: Authentication token for OpenAI access.
• ANTHROPIC_API_KEY: Authentication token for Anthropic access.
• OPENROUTER_API_KEY: Authentication token for OpenRouter proxy access.

Diagnostic Assistance

Debugging inter-process communication over stdio requires specialized utilities. We recommend utilizing the official MCP Inspector:

bash npm run inspector

This command launches an interface accessible via a browser URL, providing enhanced visibility into the communication channels.

Operational Enhancements Log

Modern Claude Model Integration (Q1 2024)

• Incorporated support for the newest Anthropic models:
  • Claude 3 Family (Haiku, Sonnet, Opus)
  • Claude 3.5 Family (Haiku, Sonnet)

Unbounded History Retrieval Feature

• Added functionality to retrieve the complete historical transcript.
• To fetch all records, invoke fetch_session_log with { "count_cap": null }.
• Specifying { "count_cap": N } retrieves the N most recent entries.
• The default limit, if unspecified, is set to ten previous interactions.