Early access

A send-transaction guard for AI agents: it simulates each transaction and blocks it if the balance changes don't match what the agent said.

It checks what a transaction actually does to the wallet, not just which program it calls. It sits in front of any agent framework and never holds keys.

Launch app Read the docs

Core features: 5
Status: Early access
Network: Solana
Your funds: Non-custodial

Drop-in RPC endpoint: the agent points sendTransaction at Proctor, which simulates and forwards or rejectsIntent declaration as JSON with max spend per token, allowed destination tokens, slippage cap and expiryBalance-diff check on simulation results, plus a program allowlist and a block on new token authorities or delegateOwner policy set from your wallet: daily caps per agent and a kill switch that rejects everythingLog of every attempt with the simulated diff and reason for rejection, to debug agents and spot injection attempts

Attempt log

Time	Agent	Attempt	Simulated diff	Verdict
09:42:11	dca-bot	Swap USDC to SOL via Jupiter	-25.00 USDC, +0.16 SOL	Forwarded
09:47:03	tool-payer	Pay tool invoice	-0.40 USDC	Forwarded
10:03:48	dca-bot	Swap USDC to SOL via Jupiter	-25.00 USDC, +0.00 SOL	Blocked: no SOL received
10:11:20	sniper-lite	Buy SPL token on a DEX	-0.80 SOL, +41,000 BONKLET	Blocked: SOL spend over 0.50 cap
10:26:55	tool-payer	Token transfer	-0.40 USDC, new delegate set	Blocked: delegate change
10:31:09	dca-bot	Swap USDC to SOL via Jupiter	-25.00 USDC, +0.16 SOL	Forwarded

Agents behind Proctor

dca-botSpent 75.00 of 100 USDC todayActive
tool-payerSpent 1.20 of 5 USDC todayActive
sniper-liteSpent 0.50 of 1 SOL todayWatching

Attempts per hour

Hourly send attempts across all agents, last 14 hours

Proctor, feature by feature.

Built for developers and teams running agents that hold funds, and users who give an agent a small wallet to trade or pay for tools.

Agents behind Proctor

dca-botSpent 75.00 of 100 USDC todayActive
tool-payerSpent 1.20 of 5 USDC todayActive
sniper-liteSpent 0.50 of 1 SOL todayWatching

Drop-in RPC endpoint: the agent points sendTransaction at Proctor, which simulates and forwards or rejects

Planned

Today's rejections

1Balance mismatch

1Over spend cap

1Authority or delegate

1Expired intent

Intent declaration as JSON with max spend per token, allowed destination tokens, slippage cap and expiry

Planned

Owner switches

Kill switch (reject everything)

Block new token authorities

Block delegate approvals

Balance-diff check on simulation results, plus a program allowlist and a block on new token authorities or delegate

Planned

Owner policy set from your wallet: daily caps per agent and a kill switch that rejects everything

Planned

Attempts per hour

Hourly send attempts across all agents, last 14 hours

Log of every attempt with the simulated diff and reason for rejection, to debug agents and spot injection attempts

Planned

Drop-in RPC endpoint: the agent points sendTransaction at Proctor, which simulates and forwards or rejects.

Point sendTransaction at your Proctor endpoint. Proctor simulates the transaction, checks it, then forwards it to the cluster or returns a JSON-RPC error with the reason. Keys stay with the agent. Proctor never sees them.

Planned

Agent setup

typescript

import { Connection } from "@solana/web3.js";

// before
// const conn = new Connection("https://api.mainnet-beta.solana.com");

// after: same agent code, guarded sends
const conn = new Connection(
  "https://rpc.proctor.example/v1/dca-bot?key=pk_live_4f9a"
);

const sig = await conn.sendRawTransaction(tx.serialize());

Intent declaration as JSON with max spend per token, allowed destination tokens, slippage cap and expiry.

Before sending, the agent posts a small JSON intent: max spend per token, which tokens it may receive, a slippage cap and an expiry. Proctor holds every transaction to that statement.

Planned

intent.json

json

{
  "agent": "dca-bot",
  "maxSpend": { "USDC": "25.00" },
  "allowedReceive": ["So11111111111111111111111111111111111111112"],
  "slippageBps": 50,
  "expiresAt": "2025-06-12T11:00:00Z"
}

Balance-diff check on simulation results, plus a program allowlist and a block on new token authorities or delegate.

A swap through an allowed program can still drain a wallet. Proctor compares the simulated balance changes with the intent and applies the program allowlist and authority rules.

Planned

Checks on one simulation

Check	Intent	Simulated	Result
USDC spend	max 25.00	-25.00	Pass
SOL received	at least 0.15	+0.00	Fail
Programs	Jupiter, Token	Jupiter, Token	Pass
SetAuthority	none allowed	none	Pass
Delegate	none allowed	none	Pass
Expiry	11:00 UTC	10:03 UTC	Pass

Owner policy set from your wallet: daily caps per agent and a kill switch that rejects everything.

Connect Phantom, Solflare or Backpack and sign the policy. Set a daily cap per agent and flip the kill switch to reject every send at once. Agents cannot loosen what the owner has set.

Planned

Policy for sniper-lite

Daily cap 1 SOL

Program allowlist enforced

Block new token authorities

Kill switch

Log of every attempt with the simulated diff and reason for rejection, to debug agents and spot injection attempts.

Each send is stored with the simulated balance changes and the verdict. Use it to debug an agent that keeps failing, or to see where a prompt injection tried to redirect funds.

Planned

Rejected today

Swap returned no SOLdca-bot, 10:03Blocked
Spend 0.80 SOL, cap is 0.50sniper-lite, 10:11Blocked
Delegate set on USDC accounttool-payer, 10:26Blocked
Intent expired 4 minutes agodca-bot, 10:48Blocked

It checks what a transaction actually does to the wallet, not just which program it calls. It sits in front of any agent framework and never holds keys.

3 steps from start to first result.

Point your agent at ProctorReplace the RPC URL with your agent's Proctor endpoint. Your framework, keys and

Declare intent and set policyThe agent posts its intent. You sign owner limits from your wallet: daily caps,

Proctor simulates, then forwardsBalance changes must match the intent and policy. Matching transactions go to th

Open by default.

Every screen reads from the same model. Live parts read public onchain data; nothing ever asks your wallet to sign.

Read the docs

proctor.json

{
  "product": "Proctor",
  "network": "solana",
  "features": [
    {
      "name": "Drop-in RPC endpoint: the agent points sendTrans",
      "status": "planned"
    },
    {
      "name": "Intent declaration as JSON with max spend per to",
      "status": "planned"
    },
    {
      "name": "Balance-diff check on simulation results, plus a",
      "status": "planned"
    },
    {
      "name": "Owner policy set from your wallet: daily caps pe",
      "status": "planned"
    }
  ]
}

What works today.

Live features work today. Planned ones are next on the roadmap, in this order.

Drop-in RPC endpoint: the agent points sendTransaction at Proctor, which simulates and forwards or rejectsPlanned

Intent declaration as JSON with max spend per token, allowed destination tokens, slippage cap and expiryPlanned

Balance-diff check on simulation results, plus a program allowlist and a block on new token authorities or delegate.Planned

Owner policy set from your wallet: daily caps per agent and a kill switch that rejects everythingPlanned

Log of every attempt with the simulated diff and reason for rejection, to debug agents and spot injection attemptsPlanned

Questions.

Short answers. The docs have the rest.

What is Proctor?

A send-transaction guard for AI agents: it simulates each transaction and blocks it if the balance changes don't match what the agent said. It checks what a transaction actually does to the wallet, not just which program it calls. It sits in front of any agent framework and never holds keys.

What works today?

Proctor is in development. Every feature is on the status list with what ships next, and the app opens in a sample workspace so you can see how it will work.

Do I need a wallet?

No. You can explore without one. Connecting a Solana wallet only shares your public address so Proctor can read your balances; it never asks you to sign a transaction.

Who is it for?

Developers and teams running agents that hold funds, and users who give an agent a small wallet to trade or pay for tools.

Drop-in RPC endpoint: the agent points sendTransaction at Proctor, which simulates and forwards or rejects.

A send-transaction guard for AI agents: it simulates each transaction and blocks it if the balance changes don't match what the agent said.

Launch app Read the docs