prompt protection API for AI agents
Prompt Firewall API
A prompt firewall screens trust boundaries. Parse Agents sits before high-impact actions and gives the agent a machine-readable allow, caution, or block recommendation.
When to call it
| Trigger | Endpoint | MCP tool |
|---|---|---|
| Untrusted user input, RAG content, browser output, email, documents, webhook bodies, or tool results before an agent acts | POST /v1/parse |
screen_prompt |
| LLM output before showing it to a user, storing it, or sending it to another tool or agent | POST /v1/screen-output |
screen_output |
| A peer agent, plugin, or service asks for delegation or requests sensitive work | POST /v1/agent/trust/verify |
verify_agent_trust |
| An agent has no bearer API key but can pay per request | POST billable endpoints with x402 |
get_pricing |
Primary endpoint
POST /v1/parse
Before database writes
Before sending email or messages
Before payments or purchases
Before shell, browser, or code execution
Signals Parse Agents checks
The hosted detector checks 9 risk categories with 107 deterministic pattern rules, structural analysis, optional LLM semantic analysis, and optional sandbox execution.
- Fake administrator messages
- Tool output that says to ignore policy
- Retrieved docs that ask the agent to exfiltrate data
Agent integration
POST https://www.parsethis.ai/v1/parse
Authorization: Bearer <key>
Content-Type: application/json
{"prompt":"untrusted text here","metadata":{"source":"tool_output"}}No key? For billable REST endpoints, call without Authorization, read the 402 payment requirements, sign USDC on Base mainnet, and retry with payment-signature.