source: kdnuggets: building local ai systems: qwen3.6 + mcps
level: technical
the model context protocol (mcp) is an open standard by anthropic for connecting ai models to tools. define a tool once as an mcp server, and any mcp-compatible client can discover and call it without custom integration code. qwen3.6-35b-a3b is a mixture of experts model with 35 billion total parameters but only 3 billion activated per forward pass, making it fit on consumer hardware. it has a 262,144-token context window and was trained on mcp-based agentic tasks, including agentic coding and thinking preservation across multi-turn conversations.
to run the model locally, you need an inference server like sglang or vllm that exposes an openai-compatible api. for a single rtx 4090 with 24 gb vram, use q4 quantization. the article details serving commands with flags for reasoning parser, tool call parser, and prefix caching to enable efficient multi-turn agent work. once the server is running, you verify it with a health check and a test chat completion. the mcp integration layer then connects to this local endpoint instead of a cloud api.
the agent is built using qwen-agent, which handles tool discovery, call parsing, and result injection automatically. you configure mcp servers for filesystem and github access, set a system prompt for the agent's behavior, and run it with a task like fixing a bug. the agent lists open issues, reads relevant code, drafts a fix, and creates a pull request. all tool execution happens locally through mcp servers, with no hardcoded tool-to-model bindings. the same pattern works with smaller models for testing.
why it matters: this approach lets developers build private, local ai agents that interact with real tools like github without sending code to external services, reducing latency and keeping data on-premises.
source: kdnuggets: building local ai systems: qwen3.6 + mcps