Design rules
- Vexthium has nothing to sign with. It runs no wallet and receives no private key or seed phrase.
- You start every state change. Each tool that would change state ends with an unsigned transaction waiting for you.
- You read it before you approve it. Each transaction gets a simulation and an explanation twice: once when it is built and again on the approval page.
- Keep little, protect it well. Prepared transactions sit encrypted for 24 hours, keys as hashes, network addresses as keyed hashes.
Keys and custody
Vexthium tools take no private key as input, and Vexthium will ask you for one at no point. A page or chat that claims to be Vexthium and asks for your seed phrase is an impostor.
send_signed_transaction is the only tool that broadcasts, and it takes a transaction already signed elsewhere. It confirms the signature targets the expected chain, then hands it to the network exactly as received.
A swap that carries the Vexthium fee pays it as a stated share of the output, collected by the Uniswap router within the swap. The quote, the preview and the approval page each show the rate, the amount and who receives it, and every figure you see is the amount left after it.
Vexthium owns no contract on Robinhood Chain. Swaps run through Uniswap's contracts, deposits through Lighter's, and any contract you deploy belongs to your wallet.
Dry run on every build
Before you see a prepared transaction, Vexthium runs it with eth_simulateV1 against the newest block of the chain it targets. The result comes back as the outcome, the gas and the movement in your own balances, written out like "−0.001 ETH, +1.78 USDG".
If the dry run reverts, the preview names the reason. The transaction is still returned, carrying a flag, so the decision stays with you, and nothing is broadcast.
Risk flags
Every transaction passes through the same fixed set of structural checks.
| What the check finds | Severity |
|---|---|
| The dry run reverts | danger |
| The sender lacks ETH for value plus gas | danger |
| The recipient is the zero address | danger |
| A token or Permit2 allowance with no limit | danger |
| Control over every NFT in a collection | danger |
| Any token or Permit2 allowance | warning |
| A Permit2 allowance lasting over 30 days | warning |
| Tokens taken from a wallet other than the sender | warning |
| Tokens sent to a contract address | warning |
| A called contract with no verified source | warning |
| Call data that matches no known ABI | warning |
| Call data aimed at an address with no code | warning |
| The dry run could not run | warning |
| Verification status could not be read | info |
| A known protocol contract without explorer verification | info |
| A contract deployment | info |
The approval link
An approval link looks like /sign#<id>.<key>. Under that id Vexthium stores the transaction, encrypted, for 24 hours. The key sits in the same fragment after the #, a part of the URL browsers keep to themselves; the page loads the stored copy by id and decrypts it inside your browser. Vexthium builds and simulates every transaction it prepares, so it sees each one in full; the encryption guards the stored copy, and the database on its own holds ciphertext with no keys. Whoever holds the complete link can open the transaction, so keep the link to yourself.
- The page checks the payload strictly and turns away anything malformed or changed.
- It runs a fresh dry run on the current block, so a link that waited in a chat for an hour is judged against the chain as it stands now.
- It asks for the connected wallet to be the sender named in the link.
- With several steps, it waits for each receipt before offering the next one.
Handle an approval link as you would any link that opens a wallet prompt: look at the address bar, and read the dry run before you approve.
Agent keys
- You create a key by signing a one-time message with your wallet, so there is no password involved at any step.
- Vexthium stores a SHA-256 hash of each key and never the key itself, which is why a key is shown only once.
- Scopes decide which tools a key may call, and the client sees only the tools inside its scopes.
- A key carries no signing power. If one leaks, the exposure is its call allowance, and your funds stay put.
- Whoever holds a key can revoke it, so a key that turns up somewhere unexpected can be switched off on the spot.
Data on our side
| Data | Kept | How |
|---|---|---|
| Seed phrases and private keys | Never | Never received |
| Prepared transactions | Encrypted, 24 hours | AES-GCM ciphertext; the key lives only in the link |
| Signed transactions | No | Passed straight on when broadcast |
| What tools receive and return | No | Left out of the log |
| A caller's network address | No | Kept as a keyed one-way hash for rate limits |
| Tool call details | Yes | Tool name, chain, outcome, error code, duration, time |
| Agent keys | Hash only | SHA-256 hash plus the first 12 characters |
| Key owner's wallet address | Yes | Used to list and revoke your keys |
| Agent profile | Yes | Name, description, capabilities; public when you choose |
The Privacy Policy covers all of this in detail.
Habits that help
- Read the explanation and each flag before you approve, and reject any step you did not ask for.
- Check the recipient and the amount in your wallet's own prompt.
- Run a new flow on the development network first by asking for chain 46630.
- Store agent keys in a secret manager, with one key per agent.
- Keep seed phrases and private keys away from assistants, websites and people alike.
Reporting a vulnerability
Send the operator a private description of the issue first, and publish details only after that. Good-faith research that leaves other people's data alone and keeps the service running is welcome, as the Terms set out.