Livy
Get Started

Introduction

How Livy protects value-moving agent actions before execution.

Livy is built for teams that want AI agents to act without giving them unrestricted wallet power.

An agent may be able to propose a transfer, payment, swap, DeFi workflow, or protocol action. Livy decides whether that action is allowed before execution and records the result afterward.

The Problem

A wallet key is not a permission model.

Agents can make useful decisions quickly, but a value-moving action still needs checks that should live outside the agent's local prompt or tool code:

  • which wallet the action is allowed to use
  • which asset and network are in scope
  • how much value can move per action or per day
  • whether the route, destination, signer, or protocol is allowed
  • whether the quote, evidence, or request context is fresh enough
  • whether the final transaction matches the approved bundle
  • whether execution produced a receipt

Livy gives those checks a backend-owned control point.

The Livy Model

Livy turns an agent action into a governed request:

agent intent -> MCP request -> wallet match -> guardrail decision -> approved execution -> receipt

The agent can ask. Livy decides. Execution only counts as protected when the approved request and final receipt are linked.

Core Objects

  • Agent wallet: the public wallet profile Livy matches against incoming MCP requests.
  • Guardrail: the policy limits and checks that decide whether a request can be approved.
  • Runtime API key: the scoped secret the agent uses to call Livy from server-side code.
  • MCP endpoint: the protected tool surface agents use to list, prepare, authorize, execute, and check actions.
  • Approval nonce: a short-lived authorization for one exact action bundle.
  • Receipt: the execution record that proves what happened after approval.

Product Boundary

Livy does not replace your agent framework, model, wallet provider, or product backend.

Livy sits at the value-moving boundary. It gives your system a place to define policy, enforce it before execution, and inspect the action trail afterward.