EX MCP Server

# Overview ## What this is This manual explains, end to end, how to install, operate, and extend the EXAI-MCP server. It is written for beginners and operators who want a reliable, repeatable way to use the system. ## Who this is for - New users evaluating EXAI-MCP - Operators running day-to-day tasks (logs, health checks, smoke tests) - Designers who want to understand the big-picture architecture and routing ## What you can do with EXAI-MCP - Use a unified tool suite over MCP (stdio/WebSocket): chat, analyze, codereview, debug, refactor, tracer, testgen, precommit, planner, thinkdeep, secaudit, consensus, status, listmodels, version, stream_demo - Route requests automatically to the most appropriate model (GLM manager by default; Kimi/Moonshot for long-context) - Validate the server with smoke tests and structured logs (MCP_CALL_SUMMARY lines show model, tokens, duration) ## How it works (at a glance) - Tools: individual capabilities exposed via MCP - Provider Registry: knows which models are available and their capabilities - IntelligentTaskRouter: chooses a model based on task cues (web, long-context, vision, etc.) - Optional WebSocket shim/daemon: lets multiple MCP clients connect easily - Logs: mcp_server.log (all activity) and mcp_activity.log (tool calls + summaries) ## Quick start Prerequisites - Python 3.9+ - Git - At least one provider key in .env (KIMI_API_KEY or GLM_API_KEY) Install ``` python -m venv .venv # Windows .venv\Scripts\activate # macOS/Linux source .venv/bin/activate pip install -r requirements.txt pip install -r requirements-dev.txt cp .env.example .env # Add KIMI_API_KEY=... and/or GLM_API_KEY=... ``` Run the server ``` python -m server ``` Verify with an MCP client: call status, version, and listmodels. ## Example end-to-end session 1) Exploration (thinkdeep) - Ask: "Assess our repo’s architecture; list risks and quick wins" - Observe model in the log summary; web-cues may route to GLM, long-context to Kimi 2) Code review (codereview) - Provide the target file(s) and focus areas; confirm issues + suggestions 3) Routing demo (stream_demo) - Run fallback (non-stream) and streaming modes to validate provider connectivity ## FAQ (short) - Why did it pick GLM when I expected Kimi? - If the prompt fit within GLM’s context window and no long-context hint was present, GLM stays preferred for speed. - How do I force long-context? - Include a large prompt and/or set an explicit estimated_tokens hint via the calling tool; see Routing. - How do I see what happened? - Check logs/mcp_activity.log for MCP_CALL_SUMMARY lines. ## Glossary - MCP: Model Context Protocol (standard interface for tools/LLM servers) - Routing hints: metadata such as long_context and estimated_tokens that bias model selection - WS shim/daemon: a small wrapper to let multiple clients use the MCP server via WebSocket