Put deterministic checks around untrusted text.
Use one preprocessing call before retrieved, pasted, or extracted text enters an agent. Then validate structured model output immediately before downstream code depends on it.
Step 1 · Prepare input
Normalize, inspect, scan, and detect secrets in one call
curl https://api.scrubmytext.com/v1/prepare-for-agent \
-H "Content-Type: application/json" \
-d '{"text":"Retrieved customer text goes here","mode":"strict","redact_sensitive":true}'The response includes agent_ready_text plus structured inspection, risk-scan, and secret-detection reports. Add Authorization: Bearer smt_live_YOUR_KEY for subscription use.
Input decision rule
Use the cleaned text only after reviewing the report
Block or route to review when risk or secret findings exceed your application's policy. ScrubMyText supplies signals; your policy decides which signals are acceptable for each task.
Step 2 · Validate output
Check structured output before code acts on it
curl https://api.scrubmytext.com/v1/validate-output \
-H "Content-Type: application/json" \
-d '{"output":"{\"action\":\"open_ticket\",\"priority\":\"high\"}","schema":{"type":"object","required":["action","priority"],"properties":{"action":{"type":"string"},"priority":{"type":"string","enum":["low","medium","high"]}}}}'Output decision rule
Pass structured output downstream only when valid
Do not execute, store, or forward the structured result unless validation succeeds. Schema validity does not authorize the requested action; apply your normal permission and business-rule checks next.
MCP workflow
Use the same guardrails from an agent
1. prepare_for_agent({"text":"UNTRUSTED_TEXT","mode":"strict","redact_sensitive":true})
2. Review the returned inspection, scan, and secret reports against policy.
3. Give only agent_ready_text to the model.
4. validate_output({"output":"MODEL_OUTPUT","schema":{"type":"object","required":["action"]}})
5. Continue only when validation succeeds and separate authorization checks pass.Limits to design around
- Security scans return deterministic and heuristic signals, not a guarantee of safety.
- Pattern-based secret detection can miss novel formats and flag harmless lookalikes.
- Normalization can change text representation; preserve source data when exact evidence matters.
- The trial is 3 premium text calls per public IP per UTC month; subscription and selected x402 routes are available after that.
