Privacy by default
- Local inference — Ollama runs on your hardware; no per-token cloud bill unless you add Claude, OpenAI-compatible APIs, or similar.
- Loopback hub — the Go hub binds to localhost; random websites cannot call your agents without crossing browser same-origin protections and your OS firewall.
- Opt-in cloud — API keys are stored encrypted; you choose which agents use which provider.
- Personal learning — memory is user-confirmed; nothing is saved without your approval.
- No telemetry page — open source; inspect the code or run air-gapped with local models only.
Human control
- File change approval — agents propose edits; you approve or reject with diff preview.
- Tool approvals — gate risky MCP and shell tools where your stack requires it.
- Collaboration phases — planning → review → approve → execute; workspace confirmation before agents touch files.
Built-in protections
| Concern | Mitigation |
|---|---|
| Cross-site calls to the hub | Loopback bind, restricted CORS, sensitive routes need loopback or hub token |
| Path traversal | pathutil.WithinRoot, resolved hub data paths |
| XSS in chat markdown | DOMPurify; unsafe style attributes disallowed |
| Secrets on disk | AES-GCM for config tokens and desktop credentials (machine-bound key) |
| API abuse | Per-IP/session rate limits (configurable) |
| Hub on LAN | Opt-in only — requires NEURAL_JUNKIE_HUB_TOKEN |
Shared or production machines
For laptops used by multiple people or servers exposed beyond localhost:
- Set
NEURAL_JUNKIE_HUB_TOKENandNEURAL_JUNKIE_CONFIG_KEYto random secrets. - Enable
NEURAL_JUNKIE_AUTH_REQUIRED=1so mutations require a session. - Do not use
NEURAL_JUNKIE_LISTEN_ALLor wildcard CORS without TLS and proxy auth. - Review MCP tools (shell, kubectl, docker) — they run as your OS user.
Desktop Settings → Security surfaces hub token, strict auth, and listen-all status. Settings → Connection (beta.5) centralizes hub URL, server/network, automation, and connector tabs without hand-editing JSON.
Lab & regulated environments
Life sciences and customer packs can add viewers and QC tools that read local scan data. Data stays on the workstation unless you configure cloud models or external integrations (Slack, Jira, AWS read-only MCP, Hugging Face for folding). Customer-specific packs are installed via sideload zip — not published to the public store.