๐Ÿ›ก๏ธ Interven
SaaS Agent Protection

Decagon

Protect Decagon AI concierge agents via API endpoint config pointing at Interven.

Decagon

Decagon's AI concierge agents call external systems via configurable API endpoints. Each endpoint has a base URL and an auth header โ€” point the URL at Interven and move the real credential into the Interven inbound route.

Setup

1. Create the inbound route in Interven

Console โ†’ Inbound Routes โ†’ Add route

  • Path prefix: decagon-account
  • Upstream URL: https://api.your-backend.example.com/v2
  • Upstream headers:
    Authorization: Bearer your-real-backend-token
    X-Internal-Auth: your-internal-key
  • Tool name: custom_proxy (or set to a built-in if matched)

2. Configure Decagon endpoint

In Decagon's agent settings โ†’ API integrations โ†’ Add endpoint:

  • URL: https://api.intervensecurity.com/inbound/decagon-account/lookup
  • Auth: none (Interven supplies the real auth on egress)
  • Request shape: match what your real backend expects

3. Wire into the conversation flow

Reference the endpoint in your Decagon flow's tool catalog. The agent invokes it during customer conversations; Interven evaluates each call before forwarding.

Behavior

Interven decisionDecagon agent sees
ALLOWReal backend response, agent continues
DENYHTTP 403 โ€” agent shows error message or escalates
SANITIZEReal backend response; redacted body went upstream
REQUIRE_APPROVALHTTP 202 โ€” agent treats as transient or hands off to human

Typical use cases

  • PII scrub before backend โ€” customer types email/phone into the chat; Interven redacts before forwarding to your CRM or billing system.
  • High-value action gating โ€” require human approval for refunds over $X, account closures, or any irreversible write.
  • Threat-intel + business-hours rules โ€” block off-hours WRITE operations on sensitive endpoints; pair with the SRE starter pack.

Caveats

  • Decagon's timeout for external calls is typically tight (5โ€“15s depending on your plan). Interven adds <100ms in normal operation but can spike on threat-intel cold lookups โ€” tune THREAT_INTEL_TIMEOUT_MS if needed.
  • One inbound route per upstream service is cleaner than a catch-all for policy attribution.