Visual Neuroscience

Ask the atlas from an AI assistant

the MCP server, and how to connect it

Visual Neuroscience answers questions inside your AI of choice, any assistant that speaks the Model Context Protocol, through a small server of its own. Ask what a region does, where a Brodmann area sits and what borders it, how dense a receptor is, or how the brain’s networks reconfigure in a state, and the answer comes from the same data as these pages, with a link that opens the same view here. Asking the atlas needs no account and no key; you sign in only if you want an assistant to read or save your own models.

The address

https://mcp.visualneuroscience.ai/

That is the whole of it: the server speaks MCP over its streamable HTTP transport at that address, with no path after the slash. Opening it in a browser lists the tools; /health answers ok.

Connecting it

The five atlas tools are read-only and say so in their annotations, so an assistant can call them without asking you to approve each one; nothing an assistant does through them changes anything, here or anywhere. Four more tools work on your own account, and ask you to sign in the first time one is reached: where an owner adds the connector for an organisation, they should set its authentication to sign in when needed.

The five tools

search_brain
Searches everything at once, AAL regions, Brodmann areas, network nodes and receptor structures, by name, number or function, and returns what each match does with a link into the atlas. “Which parts of the brain recognise faces?”
region_info
What one AAL region does, which layer of the brain it belongs to, whether receptor densities are charted for it, and a link that opens it highlighted on the scan in three planes and 3D. “What does the anterior cingulate do?”
brodmann_area
One Brodmann area by number: its anatomical name, lobe and function, the caveat that belongs with it, and which areas it borders, measured off the atlas volume. “What is Brodmann area 44, and what borders it?”
receptor_density
Receptor density in one region, or where one receptor is densest across the 45 cerebral regions, from two sources kept apart: in vivo PET, as each receptor’s share of its own densest region, and ex vivo autoradiography, in fmol/mg protein. Each answer carries its tracer, sample and citation. “Where are D2 receptors densest?” “Which receptors are densest in the hippocampus?”
brain_state
How the brain’s large-scale networks reconfigure in one of six states, flow, rest, focused attention, threat, reward pursuit and meditation: which couplings strengthen, which fall quiet, and the literature each composite is read from. Called without a state, it lists the six. “How does the brain change in the flow state?”

With your account

If you have a Visual Neuroscience account, an assistant can use it once you have allowed it to. The first time it reaches one of these four tools, the assistant shows a Connect button; that opens a page on this site where you log in as you would in the app and press Allow, and the assistant carries on where it left off. It never sees your password, and the key it is given stops working when you log out of it in the assistant, change your password or delete the account.

my_account
Who is signed in: your username, university and email. “Which account am I connected as?”
list_models
The 3D Region Atlas models saved to your account, each with what it highlights and a link that opens it. “What models have I saved?”
load_model
One saved model by name: its regions, and the link that opens it highlighted on the scan. “Open my memory circuit model.”
save_model
Saves a new model from a list of regions, which then appears in the app’s Saved models on every device signed in to the account. It only ever creates: a name already in use is refused, so an assistant can never replace or delete your work. “Save the hippocampus, the left amygdala and the thalamus as a model called memory circuit.”

To take the access away, remove the connector or disconnect it in the assistant. Nothing about the connection is kept on this server: the assistant holds the key, and the key is Firebase’s, the same service that signs you in to the app.

Where the answers come from

The server’s data is generated from the same files these pages draw, so it cannot disagree with them: the AAL-116 parcellation on the MNI152 template, the 41 Brodmann areas of the MRIcron atlas with their measured borders, receptor densities from the PET maps compiled by Hansen et al. 2022 (CC BY-NC-SA 4.0) and the autoradiography of Zilles and Palomero-Gallagher 2017 (CC BY 4.0), and the six network states with their sources. It is educational reference material, as the pages are, and not medical advice; every answer about a receptor says which tracer and sample it rests on.

What it keeps

Nothing. The server holds no accounts of its own and no state, and keeps no record of your questions beyond the short-lived operational logs of the platform it runs on, which the privacy page sets out. When you connect your account, the saved models an assistant reads or writes live in the same place the app keeps them, under your account, and the server passes them through without keeping a copy. What you type to the assistant itself is governed by that assistant’s own privacy policy.

For developers

It is one Cloudflare Worker with no dependencies, speaking JSON-RPC 2.0 over POST. Sign-in is OAuth 2.1 with PKCE, discovered the standard way (/.well-known/oauth-protected-resource and /.well-known/oauth-authorization-server), with clients registering themselves dynamically or by a client ID metadata document; a call to an account tool without a token answers 401 with the challenge that starts it. The tools can be exercised without an assistant:

curl -s https://mcp.visualneuroscience.ai/ -X POST \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Questions, corrections and requests for a tool that is missing go to fkarim@visualneuroscience.ai, and a person reads them.

Visual Neuroscience · the MCP server · OttomanLabs