This is a Model Context Protocol (MCP) server written in Python that acts as a bridge between MCP-compatible clients (like Claude Desktop) and various large language model providers. It wraps the third-party 'unichat' library to expose a single tool and a handful of code-related prompts over the MCP stdio transport.

The problem it solves is fragmentation among LLM APIs. Rather than integrating each vendor (OpenAI, Anthropic, MistralAI, xAI, Google AI, DeepSeek, Alibaba, Inception) into an MCP client separately, this server provides one MCP endpoint and lets the user pick a model and API key via environment variables. An optional UNICHAT_BASE_URL allows pointing at any OpenAI-compatible endpoint.

The intended users are developers using MCP-compatible assistants who want to relay prompts to a chosen LLM vendor. Installation is aimed at Claude Desktop configuration, uvx execution, Docker via Smithery, or local development with uv. The user supplies their own vendor API key.

How it works: on startup, server.py reads UNICHAT_MODEL, UNICHAT_API_KEY, and optional UNICHAT_BASE_URL from the environment, validates the model against unichat's built-in list (unless a custom base URL is set), and instantiates a UnifiedChatApi client. It registers a single 'unichat' tool that accepts a two-message (system + user) payload, plus four predefined prompts (code_review, document_code, explain_code, code_rework). Responses are returned as MCP TextContent. A GitHub Actions workflow auto-publishes new versions to PyPI when pyproject.toml changes.

The code is small (~10KB Python), straightforward, and shows no red flags: no telemetry, no obfuscation, no unexpected network calls beyond the LLM provider the user configures. All outbound traffic is via the user's chosen API endpoint.