Neurarch
/
Security and data handling
What we do with your keys, designs, and data
Written for the engineer doing the security review, not for a compliance checkbox. It says exactly what leaves your browser, what we store, and what we never touch. Everything here is a property of how the product is built, verifiable in the client bundle and the open repos.
Last reviewed 2026-07. For a specific question, email security@neurarch.com.
Your API keys true today
Neurarch runs LLM calls two ways, and which one fires is automatic.
- Bring your own key (BYOK). When you supply your own Claude, Gemini, Groq, OpenRouter, or xAI key, the request goes browser directly to the provider. It never transits a Neurarch server. The key lives in your browser's
localStorage, and the frontend bundle never embeds it. - Signed-in proxy. If you have not brought a key, calls route through our function with a server-side key, so we can meter quota. We store the usage event (which feature, when, token counts), never the content of your model weights or your training data.
- Key hygiene on hand-off. Signing out, or the same device switching to a different account, wipes every stored key and provider setting through a single
clearUserScopedLocalData()path. A shared machine does not leak the previous user's keys or quota.
What we store, and what we never store true today
The design decision behind the whole product: we learn from the shape of architectures, never their contents.
| We store | We never store |
|---|---|
| Structural fingerprints: a hash of the layer-type histogram and edge count, plus coarse stats (layer count, depth, parameter count). | Model weights of any kind. Trained checkpoints round-trip browser to your own Colab, Kaggle, or Hugging Face, never our storage. |
| Usage events: feature used, timestamp, quota counts, and lint verdicts (which findings you accept or dismiss). | Your training data or datasets. Dataset access happens in your own runtime (your Colab, your HF token), not on our servers. |
| Saved designs: the architecture graph as JSON (topology and hyperparameters), under row-level security so only you can read your own. | Prompt or code contents beyond what a design graph itself contains. We do not retain a transcript store of your agent conversations server-side. |
Account isolation true today
Every user-scoped artifact that lands in local storage (keys, agent memory, saved variants, edit history) is registered in one cleanup path that runs on sign-out and on any account switch. Server-side, saved designs, training records, and lint verdicts live behind Postgres row-level security policies, so a query can only ever return the caller's own rows.
The data we aggregate, done privately true today
Neurarch improves its architecture advisor from a stream of anonymous, structural signals: which lint findings users across the base trust versus treat as noise. Two safeguards make this safe to run:
- Structural only. The signal is a rule id plus a structural fingerprint, never your weights, data, or identifying design contents.
- K-anonymous crowd priors. A cross-user rule statistic is only ever returned when it aggregates at least 5 distinct users and 20 verdicts, enforced server-side. A row is always a crowd, never a person.
We do not train foundation models on your designs, and we do not sell data.
Infrastructure
| Layer | Provider | Notes |
|---|---|---|
| App and functions | Vercel | Static frontend plus serverless functions for the proxy and account APIs. |
| Database and auth | Supabase (Postgres) | Row-level security on every user-scoped table. Auth via Supabase sessions. |
| Model providers | Anthropic, Google, Groq, xAI, OpenRouter | Reached over TLS. BYOK calls never pass through us. |
| Optional training | Your Colab / Kaggle / Modal / HF | Runs in your accounts, billed to you, with your tokens. |
Enterprise and roadmap
Stated honestly: what is available now on the Team and Enterprise tiers, and what is on the roadmap and not yet certified.
Available on Team now
Shared team workspace with row-level security, per-seat access with a database-enforced seat limit, an append-only change audit log on shared designs, and design review sign-off before training.
On the roadmap planned
SOC 2 Type II attestation, SSO (Google Workspace / Okta / SAML), configurable data-retention windows, and a signed data-processing agreement (DPA). These are planned, not shipped. If your review needs one on a timeline, email us and we will tell you where it actually stands rather than claim a badge we do not hold.
Verify it yourself
- The BYOK path is client-side: open the network tab, bring your own key, and watch the request go straight to the provider host, not to neurarch.com.
- The structural fingerprint and the aggregation floors are in the open: see the public benchmark and grounding repo for how architectures are reduced to structure.