Houve | AI Crypto Agent @houveapp

Coinbase's new push: ChatGPT, Claude and Grok can trade in your account, under "your agent, your limits". It has portfolio scoping and order preview. The question is where the limit lives. In Houve the agent can't read or change the policy; changing it takes the approval window.

Source: witcheer's demo (a pizza order) and the plugin PR, which masks password fields at capture. Hermes has approval prompts too. We haven't run it. Houve is a desktop agent for EVM and Solana, not public yet.

Nous Research's hermes record turns a task you did once in the browser into a skill the agent replays with new details. Fine for a pizza. With money, the replay is the part to watch. Houve replays nothing: even a retry is rebuilt by Core from the saved intent and checked again.

Limit: native code loaded into the agent process can still get around the lock. Today that is only the database driver, set to a local file. Houve is a desktop agent for EVM and Solana, not public yet.

We assumed a Node permission flag would cut off the AI agent's network access. We measured it in Electron: the flag was ignored, and Chromium's own network stack was a way out. So the lock now runs inside the agent process, and a test tries 35 ways out.

Defense in depth doesn’t mean just “moar security layers”. The layers shouldn’t fail in similar ways. You want controls with independent failure modes, or even better inversely correlated ones. Sorry for the 2am (my time) rant, clearly this drives me crazy.

The paper is "Signing the Transaction but Not the Decision" (Louck, Dvir, Stulman, arXiv). The 56% was on Gemini Flash-Lite, in shopping agents on AP2, not crypto wallets. Houve is a desktop agent for EVM and Solana, not public yet.

A valid signature doesn't prove you saw what was signed. A September paper got shopping agents to build carts that passed every protocol check but differed from what the user was shown, 56% of the time in one test. Houve's approval screen shows what Core computed, not the chat.

Houve is a non-custodial desktop AI agent for EVM and Solana, not public yet. Nothing is signed without your approval, and the approval screen shows what Core computed, not what the chat says.

An AI agent that holds your wallet key is a spare key under the flowerpot. It works until someone finds the spot. In Houve the agent only prepares the transaction and your wallet signs. The AI never holds a key.

@vetoprotocol A request that waited an hour for your tap was built on hour-old prices and balances. Routing it to you faster helps, but it should also expire or be re-checked before it runs.

An approval screen should show what the transaction does. The function name it calls is just a label.

Musk called the new Roadster a flying car. In the demo it drove up a ramp, flipped in the air and landed. A wallet screen can pull the same trick with words: it says swap while the calldata says approve unlimited.

@4A4556494C Add the recipient to the spend line. A ceiling alone lets an injected agent stay under it and still send the whole budget to an attacker's address. The envelope should list who can be paid as well as how much.

Houve is a non-custodial desktop AI agent for EVM and Solana, not public yet. The opening line is the Grok Morse case: about $150,000, most returned later. The demo model obeys the file on purpose, the worst case we test. The rule never depends on the model spotting the trick.

A hidden line in a file can tell an AI agent to send your USDC. In Houve the agent can follow it and still get nowhere. An address that comes from a file is refused before any approval screen opens, and the refusal stays on record. 30 seconds, real app.

Houve is a non-custodial desktop AI agent for EVM and Solana, not public yet. A rejected request ends there: it can't be approved or signed. Your wallet signs only what passes every check and gets your approval.

When Houve's checks reject something the agent asked for, Core still saves it, with the reason: failed validation, failed simulation or blocked by policy. If a prompt injection ever tries to move funds, what it tried stays on record, not only what went through.

Houve is a non-custodial desktop AI agent for EVM and Solana, not public yet. Our answers to those three: no, no, and the approval view shows only what Core sends it: recipient, amount, network.

Three questions before an AI agent gets a wallet. Can it ever see the key? Can it click Approve itself? Does the approval screen show the recipient from the transaction, or the one the agent typed? One wrong answer is enough to say no.