neuroflow:setup

Agent-facing knowledge for all neuroflow integrations. Use this skill when a user asks about credentials, integration status, or setting up an external service — without necessarily running the full /setup wizard.

---

Integrations overview

Integration	Credential required	Key
PubMed / bioRxiv	❌ No	— (handled by biorxiv MCP server)
Miro	✅ Yes	MIRO_ACCESS_TOKEN
Context7	❌ No	—
Google Workspace CLI (gws)	✅ Yes	OAuth via gws auth login
Custom LLM provider	✅ Yes (API key)	custom_llm.api_key in integrations.json

---

Global vs project-level credentials

neuroflow supports two credential scopes:

Scope	Location	When to use
Global (device-wide)	~/.neuroflow/integrations.json	Credentials — set once, applies to all projects on this machine
Global user identity	~/.neuroflow/user.yaml	GitHub username + known hives — pre-fills /neuroflow setup across projects
Per-project	.neuroflow/integrations.json	Overrides global credentials for this project only

Resolution order: per-project credentials take precedence over global. If a key is not found per-project, fall back to global.

Platform paths for global config: - macOS / Linux: ~/.neuroflow/integrations.json and ~/.neuroflow/user.yaml - Windows: %USERPROFILE%\.neuroflow\integrations.json and %USERPROFILE%\.neuroflow\user.yaml

user.yaml schema:

flowie_handle: {github-username}
flowie_repo: {github-username}/flowie
hives:
  - org/repo   # one entry per team hive the user is a member of

Written by /setup Step 6. Read by /neuroflow Step 1b to pre-fill the GitHub username without asking again.

When guiding a user, ask:

"Save credentials for this project only, or globally on this machine (all projects)?"

Default recommendation: global, so they don't repeat setup on every new project.

---

How to check integration status

1. Detect platform (os.platform() or check $OSTYPE / $env:OS).
2. Read global config at ~/.neuroflow/integrations.json (or %USERPROFILE%\.neuroflow\integrations.json on Windows) if it exists.
3. Read per-project .neuroflow/integrations.json if it exists. Per-project keys override global keys.
4. Check whether the corresponding environment variables are set in the current shell (MIRO_ACCESS_TOKEN, etc.).
5. For gws: run gws --version 2>/dev/null or which gws (Unix) / where gws (Windows) to detect installation; run gws auth status to check OAuth.
6. For custom LLM: check integrations.json for a custom_llm key; if present, show provider and base_url.

Display a status table:

Integration              Status
──────────────────────   ──────
PubMed / bioRxiv         ✅ no credentials needed
Miro                     ✅ configured  (or ❌ not configured)
Context7                 ✅ no credentials needed
Google Workspace CLI     ✅ installed  (or ❌ not installed)
  └─ OAuth credentials   ✅ configured  (or ❌ not configured)
Custom LLM               ✅ configured (provider: einfra)  (or ❌ not configured)

---

integrations.json schema

Full schema including all integrations:

{
  "miro": {
    "MIRO_ACCESS_TOKEN": "eyJ..."
  },
  "google_workspace": {
    "GOOGLE_WORKSPACE_CLI_CREDENTIALS_FILE": "/home/user/.config/gws/client_secret.json",
    "GOOGLE_WORKSPACE_CLI_CLIENT_ID": "<optional>",
    "GOOGLE_WORKSPACE_CLI_CLIENT_SECRET": "<optional>"
  },
  "custom_llm": {
    "provider": "einfra",
    "base_url": "https://llm.ai.e-infra.cz/v1",
    "api_key": "<YOUR_API_KEY>",
    "model": "kimi-k2.5",
    "proxy_mode": "custom",
    "proxy_port": 4001
  }
}

• .neuroflow/integrations.json is gitignored — credentials never leave the local machine.
• When merging: only overwrite keys the user just set; leave others unchanged.
• api_key fields are always local-only, even if other non-secret fields are synced to a flowie profile.

---

Credential validation rules

Integration	Validation
Miro	Non-empty, at least 20 characters; eyJ prefix is a good sign (JWT)
Google Workspace	File must exist at platform path, or env vars must be set
Custom LLM	API key: non-empty; base URL: must start with http; model: optional

---

Custom LLM provider support

The /setup Step 5 allows configuring any OpenAI-compatible API endpoint as a replacement for Anthropic's API in Claude Code.

Model selection during setup

When setting up a custom LLM provider, always ask the user which model to use. Do not assume a default silently.

Step 1 — try to fetch available models from the API:

curl -s -H "Authorization: Bearer <API_KEY>" <BASE_URL>/models | jq '.[].id' 2>/dev/null

Or via Python:

import httpx
r = httpx.get("<BASE_URL>/models", headers={"Authorization": "Bearer <API_KEY>"})
print([m["id"] for m in r.json().get("data", [])])

If this succeeds, present the model list to the user. If it fails (auth error, endpoint not supported), skip to Step 2.

Step 2 — ask the user to pick:

I couldn't fetch the model list automatically. Please check the provider's documentation or dashboard for available model names, then tell me which one to use.

For e-INFRA CZ, the current models are listed in skills/setup/references/einfra-cc.md — but the list changes, so verify at https://llm.ai.e-infra.cz if unsure.

Step 3 — recommend based on use case:

Once you have the model list (fetched or provided by user), recommend based on their project type:

Use case	Recommend
Claude Code agentic workflows, MCP tools, neuroflow	Model with best tool-calling support (e.g. kimi-k2.5 on e-INFRA)
General research, writing, analysis	Largest general-purpose model available
Reasoning / complex multi-step tasks	Thinking/reasoning model if available
Fast iteration, simple tasks	Smallest/fastest model

Always state your recommendation and why, then confirm:

I recommend kimi-k2.5 for your setup — it has the best tool-calling performance on e-INFRA, which matters for Claude Code's agentic workflows. Use this? (Y / type a different model name)

Save the confirmed model as BIG_MODEL in the start script and as model in integrations.json.

---

Recommended approach: custom Python proxy (proxy.py)

The custom FastAPI proxy in skills/setup/scripts/einfra/proxy.py is the most reliable way to connect Claude Code to e-INFRA. It performs the full Anthropic↔OpenAI translation — streaming, multi-turn tool use, model name mapping — and has been specifically tested and fixed for the edge cases that arise with e-INFRA models.

# Terminal 1 — start proxy (Unix)
OPENAI_API_KEY=<key> OPENAI_BASE_URL=https://llm.ai.e-infra.cz/v1 \
BIG_MODEL=kimi-k2.5 SMALL_MODEL=kimi-k2.5 \
uv run --python 3.12 --with fastapi --with httpx --with uvicorn \
  uvicorn proxy:app --host 0.0.0.0 --port 4001

# Terminal 2 — connect Claude Code
ANTHROPIC_BASE_URL=http://localhost:4001 ANTHROPIC_AUTH_TOKEN=dummy claude

Windows PowerShell:

# Terminal 1 — use the provided start_proxy.ps1
.\start_proxy.ps1

# Terminal 2
$env:ANTHROPIC_BASE_URL = "http://localhost:4001"
$env:ANTHROPIC_AUTH_TOKEN = "dummy"
claude

Note: on Windows, use ANTHROPIC_AUTH_TOKEN — more reliably picked up by Claude Code CLI than ANTHROPIC_API_KEY.

⚠️ Why not LiteLLM? LiteLLM was tested with e-INFRA and produced two blocking errors: (1) Content block is not a text block with kimi-k2.5 (thinking blocks passed through incorrectly), and (2) No tool calls but found tool output with deepseek-v3.2 (tool result pairing lost in multi-turn translation). Use proxy.py instead.

⚠️ Known proxy.py fix — Content block not found: tool_block_started must be a dict mapping openai_tool_idx → assigned_block_index, not a set. Block indices assigned once at creation, never recalculated. See skills/setup/references/einfra-cc.md for details.

⚠️ Port conflicts on Windows ([WinError 10048]): Use netstat -ano | findstr :<port> to find what holds the port. Kill with cmd.exe /c "taskkill /F /PID <PID>" (must use cmd.exe, not Git Bash). Use port 4001+, avoid 4000.

For full setup, available models, Windows workflows, and troubleshooting — read skills/setup/references/einfra-cc.md.

---

Security note

• Secrets stay local. API keys and OAuth tokens are stored only in integrations.json (global or per-project), which is gitignored. Never write credentials to any other file.
• Global config (~/.neuroflow/integrations.json) is in the home directory — not inside any repo — so it can never be accidentally committed.
• Per-project config (.neuroflow/integrations.json) is gitignored via the project's .gitignore. Double-check this is in place before any git push.
• Non-secret settings (provider name, base URL, preferred model, proxy port) are safe to sync. If the user has a flowie profile linked (check ~/.neuroflow/flowie/sync.json), non-secret custom_llm settings can be written to ~/.neuroflow/flowie/integrations.json and pushed to the user's private GitHub repo for cross-machine sync. The api_key field is always excluded from this sync.

---

Running the setup wizard

The full interactive wizard is /neuroflow:setup. It covers all integrations step by step. This skill provides agent-level knowledge so Claude can guide credential setup without running the wizard — for example, answering "how do I configure Miro?" or surfacing the right reference when a user mentions e-INFRA.