Claude Code and Gemini: build a deep research agent

This workshop builds a research and writing pipeline with Claude Code, FastMCP and Gemini. The speakers explain why they give research room to explore while keeping writing in a fixed workflow. Their research system searches for evidence, analyzes YouTube videos and compiles a cited Markdown report.

The Python MCP server runs locally as a subprocess, with uv managing dependencies. Claude Code acts as the agent harness and MCP client. The walkthrough covers tool registration, prompts, resources and the Google API key configuration. Although the server runs on the user's machine, its research tools call the Gemini API, including Gemini 3 in the architecture shown. This setup depends on external services rather than offline inference.

Research results go into a .memory folder before compilation into research.md. The writing workflow combines that report with a user guideline, writing profiles and examples. A reviewer uses a separate context window and returns structured Pydantic feedback. The speakers use a fixed number of revision rounds and retain draft versions for human selection.

The final section uses Opik to inspect calls, latency and costs, then calibrates an LLM judge against human labels. The demo reports perfect F1 scores on splits containing only five samples each, but the judge later passes simulated online examples that should fail. That limitation gives viewers a concrete reason to test beyond a small evaluation set.