Your keys never leave the vault.
Engineers get a revocable token. Provider keys stay locked behind ProxDeck.
47%
of enterprise AI conversations happen through personal identities.
LayerX (opens in a new tab) · May 2026
57%
of employees say they hide their AI use and present it as their own.
KPMG / University of Melbourne (opens in a new tab) · Apr 2025
13.3%
of classified enterprise AI minutes are personal use.
Nearly half of that personal use is on enterprise plans.
Harmonic Security (opens in a new tab) · May 2026
Only 8%
of workplace AI users say their organization has a standard for the time AI saves.
AI is everywhere in most companies except the P&L.
Oliver Wyman Forum (opens in a new tab) · Oct 2026
The path a key takes
Every hop a request makes between the provider's master credential and a scoped key in the editor.
Anthropic, OpenAI and AWS Bedrock master credentials, held outside the request path and never shared downstream.
Keys are decrypted in memory for the length of a single request. No operator can read one back out.
RBAC validation, the budget breaker check and metadata dispatch, all in process with no disk I/O.
A revocable token scoped to one engineer and the branch they are on, revoked on sign-out or checkout.
Completions work normally with nothing to configure on the machine, so source never leaves the laptop for storage.
Envelope encryption
Every request is encrypted with a key generated for it alone.
Session-scoped tokens
Virtual keys expire on git checkout or screen lock, so drift cannot accumulate.
mTLS 1.3 in transit
Bidirectional verification proves packets came from an authenticated machine.
Security and cost policy in one place
Enterprise SSO & identity
GitHub SSO on every tier. Microsoft Entra ID sign-in on Pro and above. SCIM 2.0 provisioning and SAML 2.0 on Enterprise, with just-in-time provisioning and automatic de-provisioning on offboarding.
Model access policy
Model allow-lists per team, repository and environment. Model Lock goes further: it pins the models a specific feature, branch or PR may call, and the router refuses every other model rather than quietly falling back to one.
Hard budgets & circuit breakers
Stops that actually terminate the request before credit runs out. Departmental, squad and individual thresholds, with Slack and PagerDuty hooks.
Zero prompt-body retention by default
Inference runs in memory and is not persisted. We read headers to meter and route; your code never reaches storage.
What ProxDeck logs
- Timestamps & workspace IDWe log request metadata with precise timestamps, not prompt bodies.
- Model target & virtual key IDWhich service was queried, and which token authorised it.
- Token countsPrompt and completion counts, read from provider billing headers.
- Latency and PR / feature tagHow long the gateway took, plus the PR or feature it belongs to.
What ProxDeck never stores
- Prompt text & completionsNo inference payload is cached or written to logs.
- Proprietary code snippetsYour source never reaches a storage layer.
- Foundation model weightsNo training or fine-tuning on customer prompts.
- Raw upstream API keysSealed in the key vault, inaccessible to operators.
- Environment variablesNo inspection of local machine configuration.
Review the architecture with our security engineers.
Walk through the threat model and inspect the key vault isolation.