Who needs it
- Browser-agent builders that need a safe web fetch API for AI agents.
- Research and summarization agents that need clean text before downstream reasoning.
- Tool-using agents that need an SSRF guard and bounded external content handling.
Problem
Agents should not ingest raw webpage content blindly. Pages can be huge, malformed, blocked, adversarial, or filled with instructions aimed at the model rather than the human reader.
When an agent should use it
- Before summarization, browsing, research, citation gathering, or downstream tool calls.
- When the caller needs readable markdown, content hashes, and prompt-injection risk signals from a public URL.
Relevant Primitive402 tool
Fetches a public URL, extracts clean readable markdown, and flags prompt-injection patterns before an AI agent reads it.
| Primitive402 tool | safe_fetch_url |
|---|---|
| Local route | POST /v1/safe-fetch-url |
| x402 route | POST /x402/v1/safe-fetch-url |
| MCP tool name | safe_fetch_url |
| SDK method | safeFetchUrl() |
| Price | $0.01 USDC on Base Sepolia testnet during public beta |
curl example
curl -s https://primitive402.dev/v1/safe-fetch-url \
-H 'content-type: application/json' \
-d '{
"url": "https://example.com",
"output": "markdown"
}' | jq
TypeScript SDK example
import { createNano402Client } from "@nano402/api/sdk";
const client = createNano402Client({
baseUrl: "https://primitive402.dev"
});
const result = await client.safeFetchUrl({
"url": "https://example.com",
"output": "markdown"
});
console.log(result);
Example JSON response
{
"finalUrl": "https://example.com",
"cleanMarkdown": "# Example Domain\n\nThis domain is for use in examples.",
"contentHash": "sha256:...",
"riskLevel": "low",
"detectedInjectionPatterns": []
}
Useful search terms
agent-safe web fetching, safe web fetch API for AI agents, clean markdown extraction, SSRF guard, prompt-injection risk signals
Safety and limitations
Primitive402 does not guarantee content safety, factual truth, legal validity, purchase correctness, or policy completeness. External content is untrusted, low-confidence results may still be valid analysis signals, and high-stakes decisions need human review.