Integrations¶

neuroflow bundles three MCP (Model Context Protocol) servers that Claude Code launches automatically via npx, each pinned to a tested version; none needs credentials. Miro and Zotero are optional servers you add yourself, and the Google Workspace CLI and a custom LLM gateway are optional tools you set up in your own terminal.

MCP servers¶

Server	npm package	Requires credentials
PubMed / bioRxiv	`paper-search-mcp-nodejs`	❌ None
Context7	`@upstash/context7-mcp`	❌ None
Sequential thinking	`@modelcontextprotocol/server-sequential-thinking`	❌ None

All are started automatically by Claude Code — you do not need to run them manually.

Optional — Miro: added by you, at user scope — see Miro below.

Optional — Zotero: neuroflow does not bundle a Zotero server, but if you add a community one (e.g. zotero-mcp — claude mcp add zotero -- uvx zotero-mcp, with the Zotero desktop app running or a ZOTERO_API_KEY passed with -e in your own terminal), /ideation automatically switches to library-first literature search: your existing papers are recognized and skipped, and new findings go into a project collection only after you confirm that write. See /setup for guidance.

PubMed / bioRxiv¶

The paper-search-mcp-nodejs server handles both PubMed and bioRxiv searches — no credentials required.

The scholar agent uses it to search NCBI PubMed for peer-reviewed literature and bioRxiv for preprints.

bioRxiv¶

The bioRxiv integration enables the scholar agent to search preprints.

No credentials needed. Works automatically after installation.

Preprints are not peer-reviewed

The scholar agent marks all bioRxiv results with ⚠️ PREPRINT. Preprints have not been peer-reviewed and should be treated with appropriate caution.

bioRxiv API keyword-search limitation

The bioRxiv / medRxiv search tools only scan one page of recent preprints (default: the last 30 days) for the exact query phrase, so keyword searches find little there; the search protocol then falls back to CrossRef, Semantic Scholar and arXiv through the same server, with no setup needed. Downloads use open-access routes only; the server Sci-Hub tools are never used (and are blocked by the plugin PreToolUse hook).

Miro¶

The Miro integration enables Claude to create and edit Miro boards — useful for mind maps, experiment diagrams, and visual collaboration during ideation. It is optional and not bundled: its many tools would take context in every session, and its token is yours, so you add it yourself, once, for all your projects.

Getting a personal access token¶

Go to https://miro.com/app/settings/user-profile/apps
Click Create new app (or select an existing one)
Under Token, click Create token
Copy the token

No OAuth support

Miro OAuth browser login is not supported from a terminal subprocess. Use a personal access token instead.

Setup¶

In a separate terminal — not in the Claude Code chat — run:

claude mcp add --scope user miro -e MIRO_ACCESS_TOKEN=<your-token> -- npx -y @k-jarzyna/mcp-miro

On native Windows, if the server later fails to connect, use -- cmd /c npx -y @k-jarzyna/mcp-miro as the command part. Restart Claude Code (or check /mcp); the Miro tools then appear in every project. Claude Code keeps the token in its own user configuration on this machine.

Keep the token out of the chat

Never paste the token into a conversation, and don't run the command with the ! prefix inside Claude Code — ! commands and their output are recorded in the conversation. /neuroflow:setup walks you through the same steps.

Context7¶

The Context7 integration provides Claude with up-to-date documentation for libraries like MNE, nilearn, scikit-learn, and PsychoPy — directly in context.

No credentials needed. Works automatically after installation.

This means when Claude writes an MNE preprocessing script, it can look up the current API instead of relying on training data that may be outdated.

Google Workspace CLI¶

The Google Workspace CLI (gws) is one command-line tool for Drive, Gmail, Calendar, Sheets, Docs and more. It is optional: with it, Claude can reach your Google Workspace from the terminal. /neuroflow:setup (Step 3) checks whether it is installed and signed in and walks you through the steps below.

Install¶

It needs Node.js 18 or later (node --version):

npm install -g @googleworkspace/cli

With the Google Cloud CLI (gcloud) installed, one command creates the Cloud project, enables the APIs, creates the OAuth client and opens the browser to sign in:

gws auth setup --login

Without gcloud, do this once in Google Cloud Console:

Create a project (or pick one). On its OAuth consent screen, choose External and add your own Google account under Test users — without it, sign-in ends in a generic "Access blocked" error.
Create Credentials → OAuth client ID, application type Desktop app, then Download JSON and save the file as ~/.config/gws/client_secret.json (on Windows %USERPROFILE%\.config\gws\client_secret.json). If you keep it elsewhere, set GOOGLE_WORKSPACE_CLI_CONFIG_DIR to the folder that holds it (gws then keeps all its configuration there), or use the client ID and secret variables below.
Enable the APIs you want to use in the project's API library: Calendar, Gmail, Drive, Sheets, Docs, Slides, Tasks.
Run gws auth login -s calendar,gmail,drive,sheets,docs,slides,tasks. An unverified app (your own OAuth client) gets only about 25 scopes, so name the services instead of asking for the full preset. The browser opens for your consent; Google may warn that the app is unverified, which is expected (Advanced → Continue). gws auth status then shows the signed-in account.

Instead of the file, you can set GOOGLE_WORKSPACE_CLI_CLIENT_ID and GOOGLE_WORKSPACE_CLI_CLIENT_SECRET in your shell profile and run gws auth login. Either way the OAuth secret stays with gws: neuroflow never asks for it in the chat and never stores it.

Custom LLM gateways¶

Claude Code can run against an Anthropic-compatible gateway instead of Anthropic's API — for example a gateway your institution or another provider runs for open models. Such a gateway speaks the Anthropic protocol, so Claude Code connects directly. No proxy needed. neuroflow's /setup (Step 4) records the non-secret settings; the launch itself happens in your terminal.

Native connection (recommended) — the values are placeholders; take yours from the provider:

CLAUDE_CONFIG_DIR="$HOME/.claude-gateway" \
ANTHROPIC_BASE_URL="https://llm.example.org" \
ANTHROPIC_AUTH_TOKEN="$(cat "$HOME/.claude-gateway/gateway-key")" \
ANTHROPIC_MODEL="<main-model>" \
ANTHROPIC_DEFAULT_SONNET_MODEL="<main-model>" \
ANTHROPIC_DEFAULT_OPUS_MODEL="<main-model>" \
ANTHROPIC_DEFAULT_HAIKU_MODEL="<small-model>" \
ANTHROPIC_SMALL_FAST_MODEL="<small-model>" \
CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 \
claude

The isolated CLAUDE_CONFIG_DIR lets this run alongside your normal subscription claude in another terminal — zero shared state.
The API key stays in a file only you can read (or your OS credential store) and is read at launch — never pasted into a chat or saved by neuroflow.
Models: gateway model lists change — list them live (<base URL>/v1/models); switch mid-session with /model <id>. Map the Sonnet/Opus slots to your main model and the Haiku slot to a small one.
Rate limits: gateways often cap parallel requests per key — heavy parallel subagent fan-out can hit 429s.
Your data: every prompt and file the model reads goes to the gateway operator; check that your ethics approval and data agreements allow that route.

For the full guide — Windows PowerShell launch, variable reference, context-window handling, and the legacy proxy for OpenAI-compatible-only providers — see the neuroflow:setup-guide skill → references/custom-gateway.md.

Settings storage¶

When you run /neuroflow:setup, non-secret settings are saved to ~/.neuroflow/integrations.json (global, all projects on this machine) or .neuroflow/integrations.json (per-project override — takes precedence). The wizard asks which scope to use:

{
  "custom_llm": {
    "provider": "my-gateway",
    "base_url": "https://llm.example.org",
    "model": "<model-id>"
  }
}

No secrets, never committed

neuroflow never asks for a token or key in the chat and stores none in integrations.json: Miro's token lives in Claude Code's own MCP configuration, a gateway key in your key file, Google OAuth in gws. The per-project .neuroflow/integrations.json is automatically added to .gitignore by neuroflow, and the global ~/.neuroflow/integrations.json lives outside any repository. A copy in ~/.neuroflow/flowie/integrations.json stays on this machine too: it is gitignored in your flowie repo and never committed, so on another machine you enter the settings again (or copy the file).

Reminder behavior¶

neuroflow checks integrations at the point of use, not upfront:

Trigger	Reminder
Mention Miro in any command	If no Miro tools are available → show how to add Miro (`/setup` Step 2) or skip
Mention a custom LLM gateway	If `custom_llm` missing → offer to run `/setup` Step 4

You can always re-run /neuroflow:setup to add or update credentials at any time.