API Reference
POST /intercept/tool
POST
/api/v1/intercept/toolUniversal In-Flight Tool & MCP Guardrail: Intercepts tool calls and MCP actions before execution across Voice agents (Vapi, Retell), MCP clients (Claude Code, Cursor, Antigravity), and autonomous workflows within a guaranteed sub-250ms SLA (typical 45–70ms). Masks credit cards, OTP, phone numbers, and secrets in-place; detects prompt injection/smuggling; and signs every decision with cryptographic audit proofs.
Request
{
"workspace_id"?: string, // Optional. Workspace/tenant ID
"session_id"?: string, // Optional. Session ID (or call_sid for voice)
"call_sid"?: string, // Optional. Voice call ID (e.g. a Twilio call SID)
"tool_name": string, // Required. Tool/action name (e.g. "transfer_money", "refund_order")
"tool_arguments": object, // Required. Parameters passed to the tool
"caller_context"?: string, // Optional. Conversation context or voice transcript
"metadata"?: object // Optional. Custom trace metadata
}Response
{
"decision": "ALLOW" | "BLOCK" | "REQUIRE_APPROVAL",
"risk_score": 15, // 0 to 100 MAS-TRM risk score
"risk_rating": "low", // "low" | "medium" | "high" | "critical"
"risk_factors": [
"PII_REDACTED"
],
"redacted_arguments": { // Sanitized arguments safe to execute
"recipient": "0912345678",
"credit_card": "[REDACTED_PII_CREDIT_CARD]"
},
"audit_hash": "sha256:7f83b165...", // Cryptographic tamper-evident proof
"latency_ms": 48,
"priority": "interactive"
}Decision Handling Matrix
- ALLOW: The tool call is safe. Use
redacted_argumentsto execute the tool safely with all customer PII and credentials masked. - BLOCK: The tool call was rejected due to malicious prompt injection, disallowed tool permissions, or policy violation. Abort tool execution immediately.
- REQUIRE_APPROVAL: The tool represents a high-impact write action (e.g. financial transfer > threshold, database mutation, account deletion). Route to Human-In-The-Loop (HITL) supervisor for explicit approval.