Neotoma runs the same code in four places: your laptop, a tunnel pointed at your laptop, a public demo instance, and - someday - a managed production service. This page explains which hosted flavor you just landed on, or which one you want next.
Every hosted Neotoma exposes the same MCP endpoint at /mcp, the same Inspector at /app, and the same discovery endpoints (/server-info, /.well-known/*). The mode-aware root page at / adapts its copy and connect snippets to the deployment.
Hosted flavors
Public sandbox
LiveWhen to use: Evaluate Neotoma without installing. Share reproducible examples.
- Hosted by the Neotoma team at sandbox.neotoma.io.
- Data is public and resets every Sunday 00:00 UTC.
- Destructive admin endpoints disabled.
- Inspector UI is public.
Personal tunnel
Self-hostedWhen to use: You run Neotoma on your own machine and expose it to remote agents (ChatGPT, claude.ai, phone) via an HTTPS tunnel.
- Data stays on your machine; the tunnel forwards MCP traffic.
- Auth is required for writes - unauthenticated callers only see public discovery endpoints.
- Examples: ngrok, Cloudflare tunnel, Tailscale funnel, Fly.io app pointed at the local DB.
- The root URL of your tunnel renders the same mode-aware landing page shipped with Neotoma, with connect snippets prefilled to your host.
Managed production
PlannedWhen to use: Planned - you want a team-grade Neotoma with SLAs and private storage.
- Not yet offered. The self-hosted path covers most production needs today.
- Talk to us if you need a managed deployment; it informs priorities.
Local first
Most Neotoma users run locally. Your data stays on your machine, and in the default single-user setup every agent on that machine shares one dataset. You can expose it to remote agents later via a tunnel without migrating storage.
The shared-dataset default is a single-user convenience, not a limitation. Under authenticated access, Neotoma enforces per-user isolation server-side: every query is scoped to the requesting user's user_id (including graph and relationship reads), a token cannot read another user's data, and that boundary is regression-tested. One hosted instance can safely back many users.
Install guide → · Expose tunnel → · Connect a remote instance → · Try the sandbox →