@open-mercato/ai-assistant

Discovers and registers tools dynamically through the Model Context Protocol (MCP) standard, enabling AI assistants to introspect available capabilities without hardcoded tool definitions. Uses MCP's resource and tool announcement mechanisms to maintain a live registry of executable functions that can be invoked by LLM agents, supporting both local and remote tool providers.

Translates discovered MCP tool schemas into function-calling format compatible with multiple LLM providers (OpenAI, Anthropic, etc.), handling schema normalization and provider-specific function calling conventions. Manages the request-response cycle for tool invocation, including parameter validation against schemas and error handling for failed tool calls.

Maintains a conversation history that tracks both user messages and tool execution results, providing the LLM with full context about what tools have been called and their outcomes. Implements a chat loop that interleaves user input, LLM reasoning, tool invocation, and result integration, handling multi-turn conversations where tool calls may depend on previous results.

Processes raw tool execution results from MCP servers and injects them into the LLM context in a format the model can reason about. Handles different result types (JSON, text, structured data) and formats them appropriately for the LLM, managing result truncation or summarization if outputs exceed context limits.

Manages the lifecycle of MCP server connections, including initialization, health checking, and graceful shutdown. Handles both stdio-based and network-based MCP server connections, implementing reconnection logic and error recovery for transient failures. Provides connection pooling and resource cleanup to prevent leaks.

Captures and processes errors from tool execution, including schema validation failures, network errors, and tool-specific exceptions. Provides detailed diagnostic information about what failed and why, enabling the LLM to make informed decisions about retrying, using alternative tools, or reporting errors to the user. Implements structured error logging for debugging.

Provides pre-built integrations with Open Mercato-specific tools and workflows, including marketplace operations, order management, and commerce-related functions. Implements domain-specific tool schemas and execution logic tailored to Open Mercato's data models and APIs, enabling assistants to perform marketplace-specific tasks without custom tool development.

Supports streaming LLM responses while tools are being executed, enabling real-time feedback to users as the assistant reasons and acts. Implements incremental result injection where tool results become available and are streamed to the client as they complete, rather than waiting for all tools to finish before responding.