Security
Your keys, code, and traffic — handled
Cran sits in front of your LLM calls, so we hold ourselves to a simple rule: take on the credentials and the routing, not your code or your customers’ data.
Data handling
What we do — and don't — with your data
The questions every engineering lead asks before routing production traffic through a gateway.
Provider keys stay server-side
Your OpenAI / Anthropic / Google keys live on Cran, encrypted at rest — never in your repo, app bundle, Slack, or CI. Your app authenticates to Cran with a scoped connection key, not a provider key.
Your source never leaves your machine
The IDE/MCP scan runs locally and only extracts call-site metadata (file paths, models, prompt templates). Your source code is never uploaded.
Metadata, not payloads
By default the gateway records usage metadata — model, tokens, latency, cost, status — not your prompts or completions.
Capture is opt-in and redacted
Real-traffic capture is off by default and enabled per workflow. When on, samples are redacted before storage so you can benchmark on realistic inputs without exposing customer data.
Encrypted + access-controlled
Connection tokens are hashed at rest; data is encrypted in transit and at rest. Access is invite-only, with tokens scoped to read, propose, or apply.
You stay in control
Take a model offline or roll back routing in ~30s, set hard spend caps, and delete a project — and its data — anytime from the dashboard.
Governance
Controls enforced at the gateway
Budgets, guardrails, roles, and an audit trail — applied before requests reach providers, without touching your app code.
Roles & teams (RBAC)
Scope tokens to read, propose, or apply. Invite engineers per project.
Budgets that bite
Monthly caps with hard stops, per-product spend, anomaly alerts.
Guardrails
PII redaction and output-schema validation at the gateway; injection stress tests in every audit.
Audit log
Every sensitive action recorded - who changed routing, when, and from where.
Server-side keys
Provider keys live on Cran, encrypted - never in your repo, Slack, or CI.
Your prompts stay yours
We store workflow metadata - not your full repo or customer payloads.
Adopt it without handing over your code.
Start with a local, read-only scan — a map of every AI call site plus a model benchmark — before any traffic flows through Cran.