Security
Built for regulated industries.
HealthTech, FinTech, LegalTech, InsurTech — your auditors care about the same things ours do. Here's how we meet them.
The six pillars
Your data, your tenant
Customer data never leaves your network. Outbound traffic only goes to the SaaS apps your agents already touch (Gmail, Slack, etc.). On the Hardware tier, even LLM inference stays local — fully air-gappable.
Encryption at rest
OAuth tokens encrypted with AES-256-GCM using a per-tenant key. Session transcripts and config files inherit OS-level FDE. Hardware tier ships with full-disk encryption enabled out of the box. Enterprise: BYO KMS keys.
Encryption in transit
TLS 1.3 everywhere. ALB-terminated; backend connections within VPC. No plaintext on any path.
BYO OAuth apps
Use your own Google / Microsoft / Salesforce / Slack / HubSpot OAuth apps — we never see your customer credentials. PKCE flows. Refresh tokens stay encrypted in your tenant.
Scope minimization
Each integration requests the minimum OAuth scopes its agents actually use. Audit-friendly: you can see scope-by-scope which agents touch what data.
Audit log
Every action — tool call, message sent, file written — is appended to an immutable JSONL metrics log with timestamp, agent ID, and outcome. Grep-able locally; exportable to your SIEM (Datadog, Splunk, etc.) on Enterprise.
Compliance status
We share where each framework actually stands — not aspirational badges.
| Framework | Status | Notes |
|---|---|---|
| SOC 2 | In progress | Type 1 audit Q3 2026. Enterprise contracts inherit your underlying cloud or hardware controls. |
| HIPAA | Aligned | Enterprise tier ships with BAA. Use BYO KMS keys, dedicated tenant, logical separation. Not a covered entity ourselves. |
| GDPR | Aligned | You control all data residency — software runs only on infrastructure you own. Hardware tier: data never leaves your office. |
| ISO 27001 | Roadmap | On the 2027 roadmap. AWS services we run on are 27001-certified. |
Need a specific control mapped? Email dev@lopie.dev — we'll show you the file/line.
Where your model runs is your call
Three native providers — pick the one that matches your compliance posture.
Self-hosted (local)
Run models on your own hardware via Ollama. Data never leaves your network. Fully air-gappable. Zero vendor cloud dependency.
AWS Bedrock
Native Bedrock integration. Claude, Llama, and Mistral on your AWS account. HIPAA BAA from AWS. FedRAMP path available. Your data stays in your account.
OpenRouter
200+ models via API. Fastest setup for non-regulated workloads. Pay per token, no hardware commitment.
Same agents, same roles, same tools — different inference backend. Switch with one env var. See all integration paths →
The deal
We don't train on your data. Your prompts, transcripts, and customer data are not used to train any model. Period. (We don't run training infrastructure.)
We don't read your data. No analytics on prompt content. No telemetry of session bodies. Only operational metrics (counts, durations, errors) — and even those stay on your host on Self-Host.
Your data leaves only when you tell it to. If your SDR agent emails a prospect via Gmail, that prospect sees the email. That's the only outbound traffic. Internal tool calls (CRM lookups, calendar events) go through OAuth with the scopes you granted.
If we shut down, you don't lose anything. Your installation has zero phone-home dependencies. Your perpetual license remains valid. The deployment continues running indefinitely; you simply stop receiving updates.
Have a security questionnaire?
Send it over. We've answered most of them. The hard ones we'll work through with you.
Email security@inhouseagents.ai →