Security

Security

Last updated · November 10, 2026

MCPify lets an AI trade a real account, so the interesting question is not what it promises but what it is unable to do. Your wallet is the account, your balance never sits with us, and the paths that could move money out of your account do not exist in the code (Google sign-in accounts are the one, deliberate exception, described below). This page is the specific version of that claim.

Custody: who holds your balance

Your wallet is your Hyperliquid account. Your private key and seed phrase stay in your wallet and are never transmitted to us.

For wallet accounts, withdrawals have no backend endpoint anywhere in the platform: they are signed in your own wallet and broadcast from your browser. No server-side code path can move money out of your Hyperliquid account.

For wallet accounts, deposit addresses are the one place we hold funds, and only in transit. Each one is derived on demand from a server secret and your own wallet address — no deposit key is ever stored — and every use re-derives it and checks it against your address, so the only destination it can ever send to is your own account. A database record edited to point elsewhere freezes the address instead of redirecting it. Every on-chain step is journaled before it is broadcast, contract addresses are fixed in code, and conversions are checked against an independent price before anything is signed.

Google sign-in accounts are custodial by design. Their key is derived on demand from a separate server secret and your Google account ID, never stored, and checked against your account before every signature. MCPify signs only three things with it: approving your trade-only key, approving the platform fee, and a withdrawal you request. Each withdrawal needs a fresh Google sign-in (plus your two-factor code, if on), goes out once per request, and is journaled before it is sent.

The trading key is restricted by the protocol, not by our promise

Arming trading approves a Hyperliquid agent key — a key the exchange itself restricts to trading actions. It can never withdraw or transfer, regardless of what our server asks it to sign.

Server-side, a single allowlist decides what any key may ever sign: place an order, cancel an order, update leverage, or register the account under MCPify’s referral code. Anything else is refused before a signature is produced. Adding to that list is a deliberate change to the platform’s custody posture, not a configuration tweak.

Approved agent keys are stored encrypted with AES-256-GCM under a key derived through HKDF, are excluded from query results by default, and are never returned by any API or MCP tool.

Agent trading is off until you turn it on

Every connection reads only until you opt in. Order execution requires a separate, explicit switch on your account, each signed-in app carries its own trading permission, and a platform-wide kill switch sits above both.

Orders placed through the connector run under server-side caps on margin, leverage and order rate, re-read from configuration at execution time and re-checked immediately before the order is sent.

An order cannot be placed that was never quoted: place_order refuses to execute without a matching preview from the last ten minutes, keyed on the exact arguments and consumed on use.

There is no withdrawal or transfer tool in the MCP surface. Closing a position and cancelling an order are risk-reducing and deliberately not rate-limited as tightly as opening one.

Accounts are isolated structurally

No MCP tool accepts an account, address or user parameter. Every tool derives the account from the authenticated session, so there is no argument through which one user could request another user’s data. On the web API, identity comes only from the signed session token — never from a query string, body field or header.

This is not left to review: an automated suite drives two live accounts on every release and asserts that neither can read or act on the other across the REST API, the MCP surface, the admin tools and the internal service. A failure there blocks the release.

Sign-in, sessions and revocation

Sign-in is Sign-In with Ethereum (EIP-4361): the server issues a nonce, your wallet signs a message naming this site, and the server rebuilds and verifies that message itself. A failed verification burns the nonce. The site also refuses to sign a message that does not name its own origin, which is what makes the domain line real phishing protection rather than decoration.

Revocation takes effect immediately rather than when a token expires: every request re-reads your account, and signing out or deactivating invalidates the existing sessions. Disconnecting an app or revoking a personal key takes effect on its next call.

Connecting an AI app

Claude, ChatGPT and other hosts connect with OAuth 2.1: the app sends you to mcpify.trade, you sign in with your wallet, and a consent screen names the app and the address its access goes to before anything is granted. Access can only ever be delivered to callback addresses of supported AI clients — registering a new app cannot change that — and PKCE is mandatory.

What the app receives works only on the MCPify connector: it cannot be used to sign in to the website, it expires within the hour, and each renewal re-checks that the app is still connected. Disconnecting an app ends its access on the next call.

A personal key is the alternative for clients without a sign-in window. It is shown once and stored only as a hash — we cannot recover it, only replace it. Where your client supports it, send it as an Authorization header rather than in the URL; on our side the edge strips it from the logged request line and no service logs request URLs at all.

What we never collect

No email address, no phone number, no name and no identity documents — there is no field for them. No advertising or cross-site trackers. What we keep is your wallet address, your settings and the audit trail of your own orders and connector calls, which is what lets you reconstruct what an AI did on your account.

Reporting a vulnerability

If you find a security issue, report it to security@mcpify.trade before disclosing it publicly. Include enough detail to reproduce the problem. We will not pursue legal action against good-faith research that avoids privacy violations, service degradation and access to accounts that are not your own.

What this page is not

None of this makes trading safe. Perpetual futures are leveraged instruments and you can lose your entire deposit through ordinary market movement, an unwise prompt or an AI that reasons badly — none of which a security control can prevent. The caps described here are abuse guards, not a promise about outcomes.