gpt-computer-assistant

Abstracts API calls across Anthropic Claude, OpenAI GPT, and LangChain-compatible models through a unified client interface, handling provider-specific authentication, request formatting, and response parsing. Routes requests to the appropriate provider based on configuration without requiring application-level provider detection logic.

Implements the Model Context Protocol as a client that communicates with MCP servers to expose tools, resources, and prompts to LLMs. Handles MCP message serialization, request/response routing, and server lifecycle management within a Docker container, enabling standardized tool integration across different LLM providers.

Packages the LLM client, MCP integration, and orchestration logic into a Docker container that can be deployed independently of the application consuming it. Manages container lifecycle, environment variable injection for credentials, and exposes the agent via HTTP or socket interfaces, enabling infrastructure-agnostic deployment.

Integrates LangChain's agent orchestration, chain composition, and memory management capabilities to enable complex multi-step reasoning workflows. Leverages LangChain's abstractions for prompt templates, output parsing, and tool binding to reduce boilerplate when building agents that combine multiple LLM calls with external tools.

Manages API keys, model selections, and runtime parameters through environment variable injection into the Docker container. Supports provider-specific configuration (e.g., OPENAI_API_KEY, ANTHROPIC_API_KEY) and agent-level settings without requiring code changes or configuration file rebuilds.

Routes tool invocation requests from the LLM to the appropriate MCP server, executes the tool, and returns results back to the LLM for further reasoning. Handles tool schema validation, parameter marshaling, and error propagation, enabling the LLM to use external tools as part of its reasoning loop without direct knowledge of tool implementation details.

Processes streaming token sequences from LLMs and MCP tool responses, buffering and forwarding tokens to the client in real-time. Handles provider-specific streaming formats (Anthropic streaming, OpenAI streaming) and aggregates partial responses for tool invocations, enabling low-latency user feedback during agent reasoning.

Implements error handling for provider API failures, MCP server timeouts, and tool execution errors. Supports fallback to alternative providers or retry logic with exponential backoff, enabling resilient agent operation even when primary providers or tools are unavailable. Logs errors with context for debugging and monitoring.

Maintains agent state including conversation history, tool invocation results, and reasoning context across multiple interactions. Integrates with LangChain's memory abstractions (ConversationBufferMemory, etc.) to enable stateful agents that can reference previous interactions and maintain coherent multi-turn conversations without requiring the client to manage state.

Adapts provider-specific features (vision capabilities, function calling schemas, response formats) to a unified interface, translating between provider APIs while preserving capability. For example, converts OpenAI function calling schemas to Anthropic tool_use format, enabling agents to use provider-specific features without code changes when switching providers.