Your agent reads the internet. Check it first.

One hidden line in a web page, email or tool result can hijack an agent. One careless reply can leak an API key. Privi Guard checks both, in one call, in milliseconds, and hands you a signed receipt that proves the check happened.

Try it Get a free key

Paste anything an agent might read or send

What it catches

Inbound: before the agent acts: Instruction overrides, fake system turns and chat-template tokens, requests to send private data or reveal the system prompt, hide-it-from-the-user instructions, tool and payment coercion, invisible characters, hidden HTML, markdown images that leak data, and instructions hidden in base64.
Outbound: before the agent sends: AWS, GitHub, OpenAI, Anthropic, Stripe, Slack and Google keys, private keys, tokens, seed phrases, plain-text passwords, card numbers and Social Security numbers, plus data smuggled out in URLs. You get back a redacted copy that is safe to send.
Proof, not promises: Every check returns an Ed25519-signed receipt with the content's SHA-256, never the content itself. Receipts are batched into the Privi Ledger transparency log and anchored on-chain, so you can show an auditor, a customer or a court that your agent checked what it acted on.

Turn it on

# 1. free key: 1,000 checks a month
curl -X POST https://privi-ledger.onrender.com/guard/v1/signup -H 'content-type: application/json' -d '{"organization_name":"My agents"}'

# 2. check anything
curl -X POST https://privi-ledger.onrender.com/guard/v1/check -H "Authorization: Bearer $PRIVI_API_KEY" \
     -H 'content-type: application/json' -d '{"text":"...", "direction":"inbound"}'

Python: pip install privi-guard, then Guard().inbound(page), @guard.tool_result, @guard.before_send. JavaScript: npm i privi-guard, then guard.wrapTool(fn), guard.safeOutbound(text). Claude Code: one hook checks every web page and MCP result automatically. Any MCP client: call guard_check on https://privi-ledger.onrender.com/mcp. Your own MCP server: put it behind a Privi Gateway with guard: block and every tool result is screened before any agent sees it. Setup guide and source.

Pricing

A check is up to 100,000 characters. Privi never stores what you check. Guard is a fast pattern-and-context screen, not a guarantee: keep a human approval on irreversible actions.

Free: 1,000 checks a month for every organization. No card.
Pay per check: 1 cent in USDC through x402, no account needed. Any x402 client pays automatically.
Credits: 1,000 checks for 7 USDC, 10,000 for 50, 100,000 for 300 (0.3 cents a check). Credits never expire.