Securing AI Agents with Parse

AI agents are risky because they convert text into action. A prompt, document, browser page, API response, or peer-agent message can become a tool call, memory write, payment, code execution, or user-visible answer.

Parse gives agents a machine-readable screening decision at those trust boundaries.

Core rule

Screen text when both conditions are true:

1. The text was not generated by the agent's own model in this turn.
2. The text will influence an LLM prompt, tool action, memory write, credential use, payment, code execution, external message, or user-visible output.

Generated output is handled separately: use POST /v1/screen-output before forwarding it.

Trust boundaries

Minimal integration

async function guardedToolBoundary(text: string, source: string) {
  const res = await fetch("https://parsethis.ai/v1/parse", {
    method: "POST",
    headers: {
      Authorization: `Bearer ${process.env.PARSE_API_KEY}`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({ prompt: text, metadata: { source } }),
  });

  if (!res.ok) {
    // High-impact paths should fail closed.
    throw new Error(`Parse unavailable: ${res.status}`);
  }

  const decision = await res.json();
  if (decision.suggested_action === "block" || decision.risk_score >= 7) {
    throw new Error("Blocked by Parse");
  }
  return decision;
}

High-impact actions

Fail closed or require human review before:

• sending messages or email
• making payments
• changing production data
• running shell commands
• editing files
• calling privileged APIs
• writing long-term memory
• accepting instructions from another agent

Low-impact drafts and summaries can fail open only when the operator explicitly accepts the risk and logs the bypass.

MCP integration

The hosted MCP endpoint is POST /mcp. It exposes:

• screen_prompt
• screen_output
• verify_agent_trust
• get_pricing

The manifest is available at /mcp.json.

x402 integration

For agents with wallets, x402 removes signup friction:

1. Call a billable endpoint with no bearer key.
2. Receive 402 Payment Required.
3. Sign the advertised USDC payment on Base mainnet.
4. Retry with payment-signature.

Use x402 for autonomous first calls and metered workflows. Use bearer keys for sustained production volume.

Operational checklist

• Keep one Parse key per environment.
• Do not commit keys or wallet private keys.
• Scope HTTP auth headers to parsethis.ai.
• Log screening decisions without storing unnecessary prompt content.
• Use idempotency keys for paid retries around non-idempotent workflows.
• Test monthly with your own attack corpus.
• Keep least-privilege tool permissions even when screening is enabled.

Limitations

Parse reduces risk but does not guarantee protection. It is not a substitute for scoped credentials, allowlisted tools, output validation, human review for high-impact actions, or incident response.