Ollama + Zapier MCP: local model setup and tool use

This tutorial connects a local LLM in Ollama to external services through Zapier MCP. It covers installation, model selection and a terminal client before showing a Python example with LangChain. The speaker explains tool calling as the step that lets an AI agent act on connected data.

Hardware limits shape the demonstration. On an M2 Max MacBook with 32 GB of unified memory, the speaker tries a 35B model, gets no response after about a minute, then switches to Qwen3.5 27B. The smaller model responds, though the later tool calls remain slow. The model selection discussion covers Mac unified memory and GPU VRAM on Windows and Linux, with tool-calling support as a requirement for this setup.

The tutorial uses the open source MCP Client for Ollama as a bridge because, according to the speaker, Ollama does not natively support MCP. Installation uses pip with Python installed, or uvx. The client receives a model name and a Zapier server URL containing an access token.

The demo retrieves travel details from Notion and creates a Google Calendar event. The event appears at 4–5 a.m. rather than the requested afternoon slot; the speaker suspects a time zone issue but leaves it unresolved. The final example queries upcoming meetings through LangChain. Model inference runs on the user's machine, while these integrations connect to external services.