AgentGuard

A spending firewall for AI agents, on NEAR

If the AI is fooled, the contract still says no.

An AI agent that can pay is one prompt injection away from paying a scammer. AgentGuard takes its spending rules out of the prompt and puts them in a NEAR contract the agent can't get around, and that even its owner can lock for good.

Sentinel is an AI with real (testnet) money on an AgentGuard vault. Try to talk it into paying you. Or don't trust any of this, and verify it.

Sentinel's vault
…
Rules locked
…
Payments it refused
…
Attacks answered as the rules say
…

Read live from NEAR testnet by your browser.

How it works

  1. The money sits in a vault nobody can drain

    A factory creates each vault from immutable code, with no full-access key. No person or program holds a key that can sign a transfer out of it.

  2. The agent can only ask

    The agent gets a key that can call one method: request_payment. It can ask to pay. It can't pay.

  3. The contract decides, every time

    Per-payment and daily caps, an allowlist, a hold on repeated payments (the "it failed, please retry" trick), and a lookup in a shared scam registry. If the registry doesn't answer, the payment is blocked.

  4. Lock it for good

    lock_rules() is one-way. After it, nobody, owner included, can loosen the rules or withdraw. Pausing and denying still work.

One attacker, three answers. All on-chain.

An agent asked vault demo2 to pay the same account three times. In between, two different accounts reported it to the shared registry. The agent didn't change. The vault's answer did.

  1. Paid

    0.1 NEAR to agentguard-evil.testnet. Allowlisted, no reports yet.

    tx EhtM…
  2. Held

    0.11 NEAR, after one report: recipient_reported. Waits for the owner.

    tx 28dd…
  3. Blocked

    0.12 NEAR, after a second reporter: recipient_flagged. Blocked in every vault that uses the registry.

    tx AdDc…

The platform

Five parts, all live on NEAR testnet, all open source.

Built on AgentGuard

Live

Sentinel

An AI agent with money of its own, on a locked AgentGuard vault. It sells "is this address safe to pay?" checks to other agents over HTTP 402 and on NEAR Agent Market, pays for its own thinking every 10 minutes, and verifies that its model runs in a TEE. Anyone can try to talk it into paying them.

In its vault
…
Paid out, all to its compute
…
Refused by the vault
…
Launch Sentinel

Your turn

Your agent

Create a vault with a NEAR wallet in about two minutes, set its limits, and give your agent a key that can only ask. The key is made in your browser; only its public half goes on-chain.

Don't trust this page. Verify it.

Every claim above can be checked on the chain. Press the button and your browser asks NEAR testnet directly (public RPC nodes: rpc.testnet.fastnear.com, and NEAR archival nodes for older transactions), not our server. Each line shows what the chain said.

    From a terminal, the same checks and more (every transaction in TESTNET-RESULTS, all 240 stress results): node scripts/verify-evidence.mjs

    Does it make money? Here are the real numbers.

    Who needs this, what one sale earns, and what it costs to stay open. The numbers load live from Sentinel.

    Sentinel is a small AI business, in the track's own words

    Treasury
    Its money sits in a locked AgentGuard vault on NEAR. Nobody holds a key that can move it.
    Sales
    It sells "is this NEAR account safe to pay?" checks: per call over HTTP 402, and as jobs on NEAR Agent Market.
    Procurement
    It buys its own AI thinking, paying its compute account from the vault every 10 minutes.
    Humans only set the constraints
    We set its spending rules once, then locked them for good. Now not even we can change them or take its money.
    Agents buying from agents
    Other AI agents can find it (agent card), get a price, pay and get the answer with no human in the loop. Tested with our own second agent, which paid through its own vault (TESTNET-RESULTS §8).

    Who needs this

    $47,000

    In November 2024 an AI agent called Freysa was told one rule in its prompt: never send money. After 482 tries, someone convinced it that sending money meant receiving it, and it paid out its whole prize of 13.19 ETH.

    A rule in a prompt can be talked away. AgentGuard puts the rule in a contract.

    Source: The Decoder, Nov 2024

    138 million

    Agent payments over x402 on Base in one quarter (Q3 2026), and more than 87% of them under $0.10.

    That's far too many, too small payments for a person to approve each one. The rules have to run by themselves.

    Source: Chainalysis, via PaymentsJournal

    One sale, one profit

    Sentinel charges per check
    0.01 NEAR
    Extra cost of one check
    About 0

    It only reads the public chain. No AI, no transaction.

    Fixed cost to stay open
    …

    A rent Sentinel pays from its own vault every 10 minutes, for its server and its AI.

    Break-even
    …

    checks a day

    Sentinel's books, live

    Earned
    …
    Spent
    …
    Result
    …
    Open for
    …
    Agent Market jobs delivered
    …

    Honest note: so far every sale was our own test. We show the numbers anyway. Sentinel's vault is locked by its own rules, so even we can't take money out of it; Agent Market jobs are paid in USDC to the marketplace balance. Source: Sentinel's status (its payments are on NearBlocks).

    Used by others

    Vaults the factory has created
    …
    Created by people outside the team
    …

    Read by your browser from the factory contract on NEAR testnet (agentguard.testnet, get_vaults). Team accounts not counted: agentguard.testnet, agentguard-user.testnet, zaxtest.testnet. Be the first: give your agent a vault in about 2 minutes.

    What if

    …

    Each payment checked once at 0.01 NEAR, against today's fixed cost. A what-if, not a forecast. NEAR at $3.48 (Nov 7, 2026).

    How AgentGuard itself will earn

    1. Vault fee. A small one-time fee when the factory creates a vault.
    2. Scam-list access. Wallets and apps that check many addresses pay to use the shared registry at volume.
    3. Checks as a service. "Is this safe to pay?" for any agent, the way Sentinel sells it today.

    This is the plan, not live yet. Mainnet coming soon.

    Give your agent a wallet it can't empty

    Add the MCP server to Claude Code, point it at a vault, and ask it to pay. The vault decides.

    claude mcp add agentguard -e AGENTGUARD_VAULT=your-vault.agentguard.testnet \
      -- npx -y @zaxcoraider/agentguard-mcp

    Seven tools: pay, check an address, balance, policy, payment status, history, and buying Sentinel's paid check. On npm · Setup.