Turn on the on-device MCP server in Neurobase for Mac, and Claude Code, Cursor, VS Code, or Claude Desktop can drive the whole pipeline — create Neurons, add sources, train, ask with source-cited answers, generate and analyze media, and assemble and run Neurobase pipelines.
$ claude mcp add --transport http neurobase \
http://127.0.0.1:7345/mcp \
--header "Authorization: Bearer nrb_…"
✓ Added MCP server neurobase
33 tools · neurons · sources · media · pipelines · graphs
The MCP server is built into Neurobase for Mac (Mac Catalyst). It runs on-device,
binds to localhost only, and authenticates with a Bearer token — so you stay in control.
Setup
Under two minutes from a fresh install to your first call from the command line.
Open Neurobase on your Mac, go to Settings → Developer → MCP Server, and switch on Enable MCP Server. The status dot turns green and shows the live endpoint http://127.0.0.1:7345/mcp. Change the port any time if 7345 is taken.
On the same screen, tap Copy to grab the token — or a ready-made client snippet. It authenticates every request, so treat it like a password. Hit Regenerate to roll it and instantly cut off any client still using the old one.
Paste the snippet for your client below. Replace <YOUR_TOKEN> with the token you just copied — the in-app snippets already include it for you.
claude mcp add --transport http neurobase http://127.0.0.1:7345/mcp --header "Authorization: Bearer <YOUR_TOKEN>"{
"mcpServers": {
"neurobase": {
"url": "http://127.0.0.1:7345/mcp",
"headers": { "Authorization": "Bearer <YOUR_TOKEN>" }
}
}
}{
"servers": {
"neurobase": {
"type": "http",
"url": "http://127.0.0.1:7345/mcp",
"headers": { "Authorization": "Bearer <YOUR_TOKEN>" }
}
}
}Claude Desktop's config file only accepts stdio servers, so add Neurobase through the UI instead: Settings → Connectors → Add custom connector, with the URL http://127.0.0.1:7345/mcp and the header Authorization: Bearer <YOUR_TOKEN>.
Capabilities
The full Neuron-to-Neurobase pipeline plus standalone media AI, exposed as agent-first tools. Long-running work returns a job id you can poll; read-only tools are always free.
useChatHistory to continue the saved conversation.Call tools/list any time for the full machine-readable catalog with argument schemas.
Under the hood
The server is a thin adapter over the same services the app uses. Every call runs as you.
Default off, behind an explicit Settings toggle. The server binds to 127.0.0.1 — never reachable from other machines — and validates Origin and Host headers to block DNS-rebinding. Local rate limits cap call volume.
Every request carries your pairing token. Regenerate it to revoke access instantly — old clients stop the moment you roll it.
Tools call the same backend as the app, signed in as your account. Your data, quotas, and citations all behave exactly as in-app — no separate API key to manage.
Because the server lives inside the app, add_source can read a file on your Mac directly — PDFs, images, video, or audio, by path or base64 — no upload step.
Training and pipeline runs return a job id right away. Poll get_job / get_run, or use wait_for_job / wait_for_run for streamed progress — no blocking calls.
In MCP Apps hosts, get_neurobase and get_run render an interactive graph right in the chat — click nodes, run live. Elsewhere they fall back to a get_graph_image PNG and neurobase:// deep links.
Pricing
Connecting is free and works on every plan — guest, free, and paid. AI work draws from the same usage limits the app uses; read-only tools never count.
| Tools | Cost |
|---|---|
Read-only — ping, list_*, get_*, get_usage, validate_neurobase |
Always free |
| Build & edit — create, update, or delete Neurons, sources, and pipelines | Free |
AI work — add_source, train_neuron, query_neuron, run_test, run_operation, run_neurobase, generate_media, analyze_media |
Draws on your plan |
Hit a limit and the tool returns a clear error with an upgrade hint your agent can relay. Check what's left any time with get_usage.
FAQ
query_neuron returns the answer plus structured citations — source id, excerpt, page, and confidence — so an agent can quote exactly which source said what. Source attribution is first-class output, not prose.get_neurobase and get_run render an interactive graph in the chat: click a node for detail, or type an input and run the pipeline live. Other clients fall back to a get_graph_image PNG plus neurobase:// deep links.Get Neurobase for Mac, flip on the MCP server, and call it from your favorite client.
Questions? [email protected]