io.github.tlennon-ie/neurodock-mcp-cognitive-graph

5-second TL;DR

• What it does: gives Claude a memory of your work, a sense of time, a refuse-rumination guardrail, and a translator for corporate ambiguity.
• Who it's for: neurodivergent people — self-ID only, no diagnosis required, no gatekeeping.
• How to install: one command, below.

NeuroDock plugs into Claude Desktop / Claude Code / Cursor (any MCP-aware client). Local-first by default. No telemetry. AGPL-3.0-or-later.

How it fits together

You chat with Claude exactly like normal. Under the hood, Claude calls NeuroDock's local MCP servers; they read and write your local store; the result flows back into Claude's reply.

Everything runs on your laptop. Nothing leaves your machine unless you explicitly turn on a cloud option.

Install

npx --yes @neurodock/cli@latest install-all

One command. It pip-installs the five MCP servers (plus the neurodock-evals harness), wires Claude Desktop / Claude Code / Cursor, and drops a starter profile at ~/.neurodock/profile.yaml.

Then, in any conversation, try one of:

What was I working on yesterday?
Plan my morning.
Decompose this goal into atomic tasks.

Claude calls the MCP tools under the hood; you just talk.

🛑 First-time gotcha: full-quit Claude before testing

Claude only reads its MCP config at startup. Closing the window is not enough. This is the #1 silent failure people hit.

• macOS: Cmd + Q (or Claude menu → Quit Claude).
• Windows: right-click the Claude icon in the system tray → Quit.
• Linux: kill the process (pkill -f Claude or quit from the tray).

Then reopen Claude and try one of the prompts above.

Requires Python 3.11+ and Node 22+. Works on macOS, Linux, Windows.

Why @latest? If you previously ran npm install -g @neurodock/cli, npx will silently reuse that older global install instead of fetching the current version — and old versions don't have install-all. The @latest tag forces npx to resolve against the npm registry every time. Same reason we use it on update.

NeuroDock ships across two registries: the Python MCP servers (MCP = Model Context Protocol — the standard that lets Claude call local tools) on PyPI, and the user-facing CLI on npm. The CLI wraps everything else.

1. Install the six MCP servers

pip install neurodock-mcp-chronometric neurodock-mcp-cognitive-graph neurodock-mcp-task-fractionator neurodock-mcp-translation neurodock-mcp-guardrail neurodock-evals

2. Wire them into your MCP-aware client

npx --yes @neurodock/cli@latest init

Detects Claude Desktop, Claude Code, or Cursor and writes the server entries automatically.

3. Restart Claude (full quit, not close-window — see the callout above).

About the neurodock command — it lives on npm as @neurodock/cli, not on PyPI. The pip install step gives you the MCP server binaries that Claude calls over stdio; the CLI is separate. Two ways to run it:

# Option A — run via npx, no install
npx --yes @neurodock/cli@latest doctor

# Option B — install once, call 'neurodock' from anywhere
npm install -g @neurodock/cli
neurodock doctor

@neurodock/cli exposes: init, doctor, validate, update, sync, uninstall, host install, host uninstall, profile show, profile validate, install-all, install-hooks, examples, plugin add/remove/list/enable/disable/validate.

Proactive guardrails (optional, recommended)

By default NeuroDock waits for you to ask. That's the wrong shape — an ND user in hyperfocus is the one least likely to remember to run a break tool. One command flips it:

neurodock install-hooks --self-test

This wires a small Python script (bundled, stdlib-only — no extra install needed) into Claude Code's hook system. It then runs silently on every tool call and auto-fires the chronometric / rumination / sycophancy heuristics when patterns trip. You get a one-line stderr banner on the next prompt — never blocked, always dismissible.

For host-agnostic coverage (catches you working in the terminal at 02:00 too), add the standalone daemon:

neurodock install-hooks --install-daemon --self-test

That registers a per-user autostart entry (HKCU Run on Windows, LaunchAgent on macOS, systemd --user on Linux). The daemon polls every 5 min and surfaces OS-native notifications.

Opt out anytime:

neurodock install-hooks --uninstall          # removes both
export NEURODOCK_GUARDRAILS=off              # disables without removing

The browser extension carries its own equivalent (Phase 2 watchdog) — it's enabled by default and toggleable from chrome.storage.local via the popup Settings tab.

See docs/.../proactive-guardrails for the design behind this and the opt-out matrix.

Want to see it work without installing from PyPI/npm at all? TESTING_LOCAL.md walks through the from-clone path.

The install-all command above is the recommended path. The same servers are also distributed through three standard channels — all install the identical local stdio servers, so you can pick whichever fits your client:

Claude Code plugin — bundles the five MCP servers and a set of ND-aware skills in one step. Requires uv (the servers run via uvx).

/plugin marketplace add tlennon-ie/neurodock
/plugin install neurodock@neurodock

Claude Desktop extension (.mcpb) — one-click install, no config editing. One bundle per server under mcpb/; build with npx @anthropic-ai/mcpb pack mcpb/neurodock-mcp-translation and drag the .mcpb onto Claude Desktop → Settings → Extensions. Requires uv.

Official MCP Registry — each server ships a server.json and is published to registry.modelcontextprotocol.io under the io.github.tlennon-ie/* namespace, so MCP-aware clients can discover them directly.

Hosted remote server (no install) — the stateless tools (translation, guardrail, and decompose) are also served over OAuth at a hosted HTTPS endpoint, so you can use them with no local install and no uv:

https://mcp.neurodock.org/mcp

In Claude (claude.ai or Desktop) → Settings → Connectors → Add custom connector → paste that URL → complete the sign-in prompt. The eight stateless tools appear immediately. An opt-in memory surface (enable_hosted_storage, connect_byos_storage, record_fact, recall_entity, …) is also exposed but does nothing until you explicitly enable storage for your signed-in account. The personal cognitive graph and neurotype profile are never hosted — they stay on the local install. Full walkthrough: Hosted server.

The hosted remote server is live at mcp.neurodock.org.

Update

npx --yes @neurodock/cli@latest update

One command. Upgrades all six MCP servers (pip install --upgrade / uv tool install), refreshes your wired client configs, and re-registers the optional native-messaging host. Same flags as install-all (--client, --profile, --installer, --dry-run, --no-native-host, --yes).

The browser extension auto-updates through the Chrome / Firefox / Edge store; sideloaded users git pull and rebuild.

Full-quit Claude after updating so it re-reads its MCP config — same as first install.

Just want to re-shape client configs without touching package versions? Use neurodock sync.

See it in action

The cognitive substrate in a real session — you talk to Claude normally; under the hood it calls NeuroDock's MCP tools to remember, pace, and decompose.

https://github.com/user-attachments/assets/bcbf65ce-88dd-4f83-8dc1-d56bdde01bed

The browser extension — translating corporate-speak inline on Gmail, Slack, and the rest, right where you read it, without context-switching back to Claude.

https://github.com/user-attachments/assets/5fa25044-923e-4c25-95bd-f1bd63fb3d6e

Prefer a written walkthrough? Try the Claude Desktop walkthrough.

Browser extension (optional)

There's also a browser extension that translates corporate-speak inline on Gmail, Slack, Linear, Notion, GitHub, Google Docs, and Outlook. It calls the same translation tools the MCP server exposes, but surfaces them where you'd actually use them — a floating Translate button plus a right-click menu — so you don't have to context-switch back to Claude just to decode "let's circle back on this."

You pick the LLM provider. Five options: Ollama (local, default), LM Studio (local), OpenRouter (including its auto-router), Anthropic, OpenAI. The API key, if you need one, stays in chrome.storage.local and never leaves the device.

Store submission is still pending (the listing prep is done; the developer accounts and screenshots aren't). Until then, load it manually:

1. Build it: pnpm --filter @neurodock/extension-browser run build
2. Chrome / Edge: go to chrome://extensions, turn Developer mode on, click Load unpacked, and pick packages/extension-browser/.output/chrome-mv3/.
3. Firefox: go to about:debugging, click This Firefox → Load Temporary Add-on, and pick manifest.json inside packages/extension-browser/.output/firefox-mv3/.

Full per-provider setup walkthrough lives in packages/extension-browser/README.md.

What's inside

NeuroDock is built around three pillars. Each pillar is made of small, independent packages that you can use one at a time or all together.

flowchart TB
  subgraph cognitive[Cognitive substrate]
    chrono[mcp-chronometric<br/>time + sessions]
    cgraph[mcp-cognitive-graph<br/>memory]
    frac[mcp-task-fractionator<br/>decomposition]
  end
  subgraph comms[Communication layer]
    trans[mcp-translation<br/>tone + meetings]
    ext[extension-browser<br/>Gmail, Slack, ...]
  end
  subgraph safety[Clinical guardrails]
    guard[mcp-guardrail<br/>rumination / hyperfocus / sycophancy]
    clin[clinical<br/>heuristic library]
  end
  skills[skills/<br/>six markdown bundles] --> cognitive
  skills --> comms
  skills --> safety
  cli[cli/<br/>install + manage] --> cognitive
  cli --> comms
  cli --> safety

neurodock/
├── packages/
│ ├── mcp-chronometric/      Time + session + break management
│ ├── mcp-cognitive-graph/   Persistent memory + entity recall (SQLite)
│ ├── mcp-task-fractionator/ Decompose vague goals into atomic tasks
│ ├── mcp-translation/       Corporate-speak translator (MCP + browser ext)
│ ├── mcp-guardrail/         Rumination / hyperfocus / sycophancy detectors
│ ├── skills/                Six SKILL.md bundles activating on phrases
│ ├── extension-browser/     WXT-built Chrome / Firefox / Edge extension
│ ├── native-host/           Optional native messaging host (extension <-> profile.yaml)
│ ├── cli/                   `npx neurodock init` and friends
│ ├── core/                  Shared types, profile schema, plugin spec
│ ├── clinical/              Heuristic library for the guardrail server
│ └── evals/                 Eval harness + corpus contribution pipeline
├── docs/                    Astro Starlight site (deploys to docs.neurodock.org)
├── plugins/                 Drop your own plugins here; auto-discovered
└── profiles/                Curated profile presets

Status

Public preview shipped. All three substrate pillars (cognitive, communication, guardrails) are built, on main, and installable from npm + PyPI. Latest substrate tag: v0.7.3.

MCP servers (PyPI)

• neurodock-mcp-chronometric — 5 tools, 22 tests, mypy --strict.
• neurodock-mcp-cognitive-graph — 4 tools, SQLite + sqlite-vec
  • fastembed; 4-rung resolution cascade (exact → alias → fuzzy → embedding). Latest patch ships friendlier record_fact errors so wrong-shape input no longer leaks raw Pydantic traces.
• neurodock-mcp-task-fractionator — 2 tools, 32 tests; ISO 8601 duration spec.
• neurodock-mcp-translation — 4 tools, 29 tests; deterministic baseline plus LLM refinement envelope.
• neurodock-mcp-guardrail — all three detectors live: rumination, hyperfocus, sycophancy (48 tests, public heuristics).
• neurodock-evals — air-gapped harness, 10 seed corpus examples, contribution pipeline.
• neurodock-clinical — reserved name; importable detector library (currently a stub).

CLI + browser surface (npm)

• @neurodock/cli — 19 verbs across 7 groups: init, install-all, examples, doctor, validate, update, sync, uninstall, host install/uninstall, profile show/validate, install-hooks, plugin add/remove/list/enable/disable/validate. install-hooks wires the proactive-guardrail hook into Claude Code and optionally registers the standalone daemon.
• @neurodock/core — profile schema and plugin protocol manifests (JSON Schema 2020-12).
• @neurodock/native-host — optional Chrome Native Messaging host for the extension ↔ profile sync.
• @neurodock/extension-browser — WXT MV3 build for Chrome / Firefox / Edge. Seven sites wired. Five real LLM providers: Ollama, LM Studio, Anthropic, OpenAI, OpenRouter. Includes the Phase 2 proactive watchdog for hyperfocus / late-night / single-host rumination signals. Not yet store-published.

Skills + docs

• Six launch skills — adhd-daily-planner, audhd-context-recovery, ocd-decision-finalizer (beta), hyperfocus-formatter, visual-organizer, asd-meeting-translator.
• Docs site — builds clean (Astro Starlight), deploys to docs.neurodock.org.

Still deferred

• Browser-store submissions — Chrome Web Store, Firefox Add-ons, Edge Add-ons developer accounts + screenshots; manual step.

How to actually test it right now

TESTING_LOCAL.md — step-by-step guide to running this against your Claude Desktop, today, from a clone. Takes about 5 minutes.

Documentation

The docs site lives at docs.neurodock.org. Source is at docs/src/content/docs/. To preview locally:

pnpm --filter @neurodock/docs run dev
# opens http://localhost:4321

Architecture

The substrate splits into three pillars:

1. Cognitive substrate — externalises executive function (time, memory, decomposition). MCP servers: chronometric, cognitive-graph, task-fractionator.
2. Communication layer — translates corporate ambiguity; rewrites outgoing messages for register-appropriate tone; structures meeting transcripts. MCP server: translation. Browser extension surfaces the same prompts in Gmail / Slack / Linear / Notion / GitHub / Docs / Outlook.
3. Clinical guardrails — detects and intervenes on rumination (repeat-validation loops), hyperfocus (escalating session-length nudges), and sycophancy (unconditional agreement). MCP server: guardrail. Heuristics are public and auditable per ETHICS.md.

All three layers compose via the same MCP protocol the LLM client already speaks. There's no "NeuroDock app" — the surface is your Claude client.

Design rationale lives in docs/decisions/:

• ADR 0001 — chronometric tool design
• ADR 0002 — cognitive-graph tool design
• ADR 0003 — task-fractionator tool design
• ADR 0004 — profile schema
• ADR 0005 — translation tool design
• ADR 0006 — guardrail tool design
• ADR 0007 — plugin protocol
• ADR 0008 — distribution & remote strategy

Contributing

CONTRIBUTING.md has the welcome + on-ramp. Pick whichever lane matches what you want to do:

Smallest first PR (~15 min):

• Add a test to an existing skill (packages/skills/<name>/tests/)
• Add a seed eval example (packages/evals/corpora/translation/)
• Improve a tool's parameter description (the LLM uses these — clearer descriptions = better tool use)

One-afternoon PR:

• Write a new skill — markdown bundle activating on specific phrases. Copy any packages/skills/<name>/ as a template; see docs/src/content/docs/contribute/write-a-skill.mdx.
• Sharpen the task-fractionator heuristics (regex / domain keyword maps).
• Add a new entity-resolution heuristic (e.g., phonetic) to mcp-cognitive-graph.

Multi-day PR:

• Build an out-of-tree plugin (skill / mcp-server / profile / translation pack / language pack / theme) per the ADR 0007 plugin protocol. See docs/src/content/docs/contribute/write-a-plugin.mdx.
• Translation language pack (e.g., Hiberno-English, German directness norms, Japanese keigo).

No code:

• Add an anonymised eval example from your own corporate inbox. The contribution pipeline lives in packages/evals/ and is the highest- leverage non-code contribution.

All PRs run CI: pnpm turbo run lint typecheck test build + uv run pytest

• uv run mypy --strict packages/.... Every published package has its own CHANGELOG. Use Conventional Commits in PR titles (feat:, fix:, docs:, test:, chore:, ci:).

Manifesto (short)

1. Lower friction for users, and for contributors.
2. Local-first by default; cloud is opt-in.
3. The user is the authority. Self-ID sufficient.
4. Composable over monolithic. No god-modules.
5. Refuse where appropriate. AI that fuels rumination, hyperfocus, or anxiety is a regression, not a feature.

Full text in MANIFESTO.md. Ethics framework in ETHICS.md. Governance in GOVERNANCE.md.

License

AGPL-3.0-or-later. Plugins must declare an AGPL-compatible license to load — the SPDX whitelist is in ADR 0007.