You don't need to write any code to work through this workshop. Two natural-language routes reach the same VFB knowledge base the Python notebooks use:
-
chat.virtualflybrain.org — a guardrailed chat interface. Nothing to install, nothing to configure. Best starting point, and ideal for teaching. →
chat_vfb_guide.md -
VFB MCP tool (vfb3-mcp.virtualflybrain.org) — connect VFB to your own LLM client (Claude Desktop/Code, Copilot, VS Code…) so the model can query VFB directly and ground its answers in real data. →
MCP_setup.md
Then work through the shared problems in ../problems/ using
example_prompts.md — every problem lists a B (MCP) and C (chat) prompt
next to the Python code, so you can compare routes.
The VFB MCP tool was benchmarked on 30 neuroscience tasks: an LLM with the MCP answered 25/30 correctly vs 14/30 for a web-search-assisted LLM and 2/30 for a bare LLM — and on tasks needing data quantification, 89% vs 11%. The gain comes from grounding answers in VFB's expert-curated, ontology-backed knowledge graph rather than the model's memory. See McLachlan et al. (2026).
Reproducibility note for researchers: natural-language routes are excellent for exploration and for getting to the right query fast, but for anything you'll publish, capture the underlying IDs and re-run the equivalent
vfb_connectcall (route A) so your analysis is scriptable and versioned.