Setup | VFB Self-Led Learning

Two of the five routes need zero setup (Chat and the 3D Browser); Python and R each need a single install command, and MCP needs you to point your own AI assistant at a URL. Do the ones you plan to use before you start; each takes a few minutes at most.

Route A — Python (`vfb_connect`)

Zero-install option: run the notebooks in Google Colab — nothing on your machine: Open 00_Setup_and_Orientation in Colab . In Colab, run the first pip install cell and, if prompted, restart the runtime before continuing.

Local install (Python 3.9+, a virtual environment keeps things clean):

python3 -m venv vfb-env
source vfb-env/bin/activate        # Windows: vfb-env\Scripts\activate
pip install --upgrade pip setuptools
pip install vfb-connect

Verify it works:

from vfb_connect import vfb        # first import connects to VFB services
df = vfb.get_instances("DA1 lPN", return_dataframe=True)
print(len(df))                     # -> 68

The first import establishes connections and caches term data, so give it a couple of minutes the first time; after that it is fast. navis (for 3D plotting) is installed as a dependency — .plot3d() works out of the box in Jupyter/Colab.

Route B — R (via reticulate)

R users get the identical vfb_connect engine through reticulate — same calls, same numbers, ordinary R data frames back.

In Google Colab: Runtime → Change runtime type → R (or create an R notebook at colab.research.google.com/#create=true&language=r), then in the first cell:

install.packages("reticulate")
reticulate::py_install("vfb-connect", pip = TRUE)

Locally (R 4.x):

install.packages("reticulate")          # use current CRAN reticulate —
                                        # older versions cannot convert pandas 3 data frames
reticulate::py_install("vfb-connect", pip = TRUE)

Verify it works:

library(reticulate)
vfb <- import("vfb_connect")$vfb        # first call connects, ~2 min
df <- vfb$get_instances("DA1 lPN", return_dataframe = TRUE)
nrow(df)                                # -> 68

Two R-specific habits: write integer arguments with an L suffix (weight = 20L), and flatten list-columns with sapply(col, paste, collapse=",") before table()/dplyr. natverse users: natverse/vfbconnectr wraps the same package and adds read.neurons.vfb() to pull skeletons straight into nat (verified working against vfb-connect 2.4.2).

Route C — Your LLM + the VFB MCP

The Model Context Protocol (MCP) lets your own AI assistant call VFB directly, so it answers from live VFB data instead of memory. The hosted server is:

https://vfb3-mcp.virtualflybrain.org

Nothing to install — you register that URL with your client. Source: VirtualFlyBrain/VFB3-MCP .

Claude Desktop: Settings → Connectors → Add custom connector — name it virtual-fly-brain, type HTTP, paste the URL.

Claude Code: add to ~/.claude.json (Windows: %USERPROFILE%\.claude.json) and restart:

{
  "mcpServers": {
    "virtual-fly-brain": {
      "type": "http",
      "url": "https://vfb3-mcp.virtualflybrain.org",
      "tools": ["*"]
    }
  }
}

VS Code / GitHub Copilot: add an MCP server in Settings (search “MCP”), or in mcp.json:

{ "servers": { "virtual-fly-brain": { "type": "http", "url": "https://vfb3-mcp.virtualflybrain.org" } } }

Any other MCP-capable client works with the same URL. (Gemini’s web UI has no direct MCP support yet — use a small Python/Node MCP client instead.)

Verify it works: ask your assistant “What is DA1 lPN? Use the VFB tools.” — you should see it call search_terms / get_term_info and answer with the FBbt ID and literature sources. If it answers instantly with no tool calls, it is answering from memory: check the connector is enabled for the conversation.

Route D — VFB Chat

Nothing to set up. Open chat.virtualflybrain.org in any browser and ask a question.

Worth knowing before you start:

You get 100 queries per day (the counter sits next to the input box).
Short, single-entity questions resolve best — “Which neurons are downstream of DA1_lPN_R?” beats a long multi-clause sentence. Reuse the names the chat itself uses in its answers.
You can pre-load a question in the URL: https://chat.virtualflybrain.org/?query=What+is+DA1+lPN%3F — this is how the embedded examples on the session pages work.
Answers are AI-generated from VFB data: verify anything critical against the primary sources it links.

Route E — 3D Circuit Browser

Nothing to set up. Open v2.virtualflybrain.org in a modern browser (WebGL required — any recent Chrome, Firefox, Safari or Edge; the first load takes a few seconds while the template brain streams in).