Ollama MCP Server

# Active Context ## Current Work Focus - Modernizing the MCP Ollama server to use the latest @modelcontextprotocol/sdk and ollama-js. - Ensuring all core Ollama operations are registered as MCP tools using registerTool and Zod validation. - Maintaining compatibility with MCP conventions and Ollama updates. ## Recent Changes - Migrated from legacy MCP server API to the modern McpServer/registerTool pattern. - Replaced class-based and CLI/axios architecture with functional handlers and ollama-js. - All core Ollama operations mapped to MCP tools with type-safe handlers. - Only stdio transport is currently implemented; HTTP/SSE is planned for future releases. - Updated documentation and memory bank to reflect new architecture. - Simplified "run" tool: removed streaming logic, now always returns full response after completion (no partial/streamed output in stdio mode). ## Recent Changes - Added support for images in chat messages (`chat_completion` tool) for vision/multimodal models. - Updated README with usage examples for chat with images. ## Next Steps - Expand test coverage for all MCP tools, including chat with images. - Monitor for changes in Ollama or MCP SDK APIs and update handlers as needed. - Add HTTP/SSE transport for remote/multi-client access. - Improve error reporting and diagnostics. - Document advanced usage and troubleshooting. ### Release Flow - Push changes and tags: `git push origin main --follow-tags` ## Active Decisions and Considerations - Using functional registerTool pattern for extensibility and clarity. - Prioritizing full API coverage and OpenAI compatibility. - Keeping configuration simple but flexible (env vars for endpoint, timeout, temperature). - "run" tool now supports an optional temperature parameter directly in its arguments for deterministic or creative outputs. - "run" tool does not support streaming in stdio MCP mode; response is accumulated and returned only after completion. - Focusing on robust error handling and compatibility with MCP SDK. ## Important Patterns and Preferences - Tool registration via registerTool with Zod schemas for validation. - Functional handler pattern for mapping MCP tools to Ollama operations. - Preference for explicit, well-documented interfaces. ## Learnings and Project Insights - MCP tool abstraction enables rapid integration of new Ollama features. - Streaming is not supported in stdio MCP transport; future HTTP/SSE support may enable true streaming. - Clear documentation and memory bank are essential for maintainability. - Modern MCP SDK and ollama-js simplify integration and improve type safety. ## Recommended Follow-ups - All contributions must follow the Conventional Commits standard (see project cline rules). - Integrate semantic-release to automate versioning and npm publishing based on commit messages. - Enable CodeQL or Snyk for automated vulnerability scanning of dependencies and code. - Activate GitHub Discussions to provide a space for community Q&A and support.