StackIntercept is designed for self-hosted deployment inside your own network path.

Client app -> StackIntercept -> LLM provider

StackIntercept does not send prompts, responses, or metrics to any external telemetry service. Admin metrics are exposed only through the local /admin/* HTTP routes.

• Provider API keys are accepted through request Authorization headers or environment variables.
• Admin API access should use STACK_INTERCEPT_ADMIN_KEY outside local loopback development.
• Do not commit .env, provider keys, cache snapshots, or model provider credentials.
• Prefer a secret manager for production deployments.

For production or shared staging:

• keep StackIntercept behind private networking or TLS termination
• do not expose port 8080 directly to the public internet
• set STACK_INTERCEPT_ADMIN_KEY
• restrict admin routes at the ingress/firewall layer when possible

Exact cache keys include provider, tenant, routing namespace, and canonical request payload.

Semantic cache is opt-in. It hashes stable conversation context first and only scans within the matching context bucket. Use this header when semantic reuse is not acceptable for a request:

x-stack-intercept-no-semantic-cache: true

Use this header when model routing is not acceptable:

x-stack-intercept-no-route: true

If STACK_INTERCEPT_CACHE_PATH is set, exact and semantic cache snapshots are written to disk using MessagePack. Treat these files as sensitive because they may contain cached provider responses.

Recommended:

• store snapshots on encrypted disks
• restrict file permissions to the service user
• disable persistence for highly sensitive trials:

STACK_INTERCEPT_DISABLE_PERSISTENCE=true

StackIntercept is not an identity provider, WAF, billing engine, or full API gateway. It currently does not enforce per-user auth, hard token budgets, request signing, or tenant rate limits.