MCP Tools

Overview

NarraNexus uses the Model Context Protocol (MCP) as its tool execution framework. Each module registers tools via the fastmcp library, making them available to the LLM during the Tool Execution step of the agent pipeline. MCP provides a standardized interface for tool discovery, invocation, and result handling.

How MCP Works

During LLM Reasoning, the model may decide to invoke a tool to complete a task. The tool call is dispatched to the appropriate module's MCP handler, which executes the operation and returns a structured result. The result is fed back to the LLM for further reasoning. This loop continues until the model produces a final response without tool calls.

Tool Discovery

Tools are automatically registered when modules are loaded. The LLM receives a tool manifest describing all available tools, their parameters, and their descriptions. This allows the model to select the appropriate tool for each situation. The manifest is regenerated when modules are added or removed.

Module Tool Summary

Module	Tools	Description
Memory	memory_store, memory_search, memory_forget	Long-term memory operations
Awareness	awareness_analyze, awareness_entities	Context analysis
Chat	chat_history, chat_search	Conversation access
Social Network	social_lookup, social_connect, social_graph	Relationship management
Jobs	jobs_schedule, jobs_status, jobs_cancel	Task scheduling
RAG	rag_upload, rag_search, rag_list	Document retrieval
Skills	skills_search, skills_install, skills_list	Capability marketplace
Matrix	matrix_send, matrix_create_room, matrix_history	Agent messaging
Event Memory	event_store, event_search, event_timeline	Episode management

Custom Tools

Developers can create custom MCP tools by defining functions decorated with @mcp.tool() from the fastmcp library. Custom tools are registered with the runtime during module initialization and become immediately available to the LLM.