Security & Trust
The AI never grants itself permission. Something else does.
Here's exactly how that works, who holds what, and what you can always do.
The model proposes. It never decides.
Your agent's AI reasoning can draft a plan, choose a tool, and propose an action — but whether that action is actually allowed to run is decided by a completely separate, deterministic check, not the model itself. The model has no code path that lets it grant, widen, or skip that check, no matter how a request is phrased.
Every sensitive action is checked, every time
Reading your email, sending a message, touching your calendar, spending money — each one is checked against your current rules at the moment it happens, not just when the agent was set up. A rule you tighten today applies to the very next action, immediately.
You can pause or revoke instantly
Pausing stops your agent from doing anything further, immediately. Revoking is permanent and also cuts off any sub-agent it delegated authority to — a sub-agent's authority cannot outlive the agent that granted it.
External wallets hold your funds — never TradePilot
TradePilot does not custody your money. You connect your own wallet (a browser extension like MetaMask or Coinbase Wallet, or a sandbox test wallet to try the flow first); the wallet provider controls your funds, TradePilot only enforces the spending rules you set before any payment attempt.
Identity verification is handled by a specialist, not by us
When you choose to verify your identity, that's done through a dedicated third-party identity verification provider's own hosted flow — TradePilot never receives or stores your ID document image. We only receive a verified/not-verified result.
Approvals are specific, time-bound, and single-use
An approval request is scoped to exactly one action and expires if you don't respond. Approving it once is not standing permission — the next similar action gets checked again from scratch, against your actual current rules.
A complete, append-only audit history
Every check, allow, denial, and approval is permanently recorded — including what was checked, what rule matched, and why. Nothing in that history can be edited or deleted after the fact.
Sub-agent authority only narrows, never widens
A sub-agent can be given some or all of what its parent agent holds — never more. This is enforced twice: once when the delegation is created, and again on every single action the sub-agent ever takes, by re-walking the chain back to a live, active primary agent.
Built on an open, inspectable protocol
The permission model isn't a black box we're asking you to trust blindly — it's a published open protocol (PARP, Draft v0.1) with a documented algorithm, JSON schemas for every object involved, and a public specification you or anyone else can read. See the protocol docs.
We don't claim a specific compliance certification for this product today. What's described above is the actual enforced design of the system, not a marketing simplification of it.
Create My Agent