Skip to content
NeuroDock

Hosted server (remote MCP)

NeuroDock runs fully on your machine by default (see Installation). But the stateless tools — communication translation, the guardrail checks, and task decomposition — are also served from a hosted endpoint you can add to Claude with no local install and no uv.

https://mcp.neurodock.org/mcp

Transport is Streamable HTTP; auth is OAuth 2.1 (PKCE). The personal cognitive graph and the neurotype profile are never hosted — they stay on the local install. The hosted endpoint exposes only the stateless surface, plus an opt-in memory surface that does nothing until you turn it on.

SurfaceHosted?Tools
Translationyes (stateless)translate_incoming, check_tone, rewrite_outgoing, brief_meeting
Guardrailyes (stateless)check_rumination, check_hyperfocus, check_sycophancy
Task fractionatorpartialdecompose only (next_one reads the local graph, so it stays local)
Memory (opt-in)only if you enable itenable_hosted_storage, connect_byos_storage, disable_and_erase_storage, disconnect_storage, storage_status, recall_entity, record_fact, recall_decisions, weekly_rollup
Chronometric, profilenonever hosted — local install only
  1. In Claude (claude.ai or Claude Desktop), open Settings → Connectors.
  2. Choose Add custom connector.
  3. Paste the endpoint: https://mcp.neurodock.org/mcp.
  4. Complete the sign-in prompt in the OAuth window. After that, the eight stateless tools are available in any chat.

That’s it for the stateless tools — they need no further setup and store nothing.

You don’t call the tools by name; you ask Claude in plain language and it routes to the matching tool. For example:

  • “Translate this Slack message — what are they actually asking me to do?” (then paste the message) → translate_incoming
  • “Check the tone of this draft before I send it” (then paste the draft) → check_tone
  • “Rewrite this so it lands as neutral and direct” (then paste the draft) → rewrite_outgoing
  • “Break this goal into startable steps: ship the onboarding revamp”decompose
  • “Is this prompt a guardrail concern?”check_rumination / check_hyperfocus / check_sycophancy

Each returns a deterministic baseline plus a structured prompt your client’s model refines — no model SDK runs on the server (vendor-neutral by ADR 0005).

The memory tools are visible but inert until you choose a storage mode for your signed-in account. You are always in exactly one mode:

  • enable_hosted_storage — NeuroDock provisions a private, per-account database and stores your cognitive-graph facts there. Returns a consent disclosure. Quickest path.
  • connect_byos_storage — point it at your own libSQL/Turso database; NeuroDock keeps only the connection pointer. Maximum privacy — the data lives in your infrastructure.
  • storage_status — reports whether you’re signed in and which mode is active (hosted | byos | none).
  • disable_and_erase_storage — hosted: destroys the database; BYOS: clears the stored connection. Either way your stored preference is cleared.

Until you run one of the enable tools, storage_status reports mode: none and the memory tools refuse with a structured “not connected” message — anonymous and not-opted-in sessions stay completely stateless. A signed-out caller is refused before anything is stored, read, or provisioned.

The hosted server processes tool inputs in memory to produce the analysis and persists nothing server-side for the stateless tools. The opt-in memory surface stores only what you explicitly enable, isolated to your account, never aggregated, and erasable at any time. See the privacy policy and ethics for the full position; ADR 0009 records the transport, auth, and hosting decisions.