SafeGPT
SafeGPT is an AI-powered chat assistant helping users interact with the Safe Protocol to set up and manage smart multi-sig accounts for secure and efficient onchain asset management.
Created on 11th February 2025
•
SafeGPT
SafeGPT is an AI-powered chat assistant helping users interact with the Safe Protocol to set up and manage smart multi-sig accounts for secure and efficient onchain asset management.
The problem SafeGPT solves
SafeGPT is an AI-powered assistant designed to simplify and secure interactions with the Safe Protocol. Safe smart accounts enhance security through multi-signature (multi-sig) capabilities, requiring multiple approvals for transactions to protect against private key leaks, the leading cause of crypto hacks. However, setting up and managing these accounts can be complex, especially for users unfamiliar with smart contracts and onchain transactions.
With SafeGPT, anyone can create and manage Safe multi-sig wallets effortlessly, without needing to write code or perform onchain transactions themselves. The AI-powered agent assists users in setting up accounts, managing signers as well as proposing, approving, and executing transactions in a frictionless way.
Beyond simplifying account management, SafeGPT enhances security by performing risk analysis on pending transactions. It detects changes to Safe wallet settings, flags high-value transfers, checks the onchain reputation of transaction recipients, checks if smart contracts are verified on Etherscan and identifies risky token transfers that could result in fund loss.
Whether users want to experiment with multi-sig wallets in a risk-free environment, analyze pending transactions before signing them, or set up automated spending limits for agents, SafeGPT makes onchain asset management both easier and significantly safer.
Challenges I ran into
Integrating the Safe SDK into AgentKit while ensuring all core functionality worked smoothly was a significant challenge. One of the toughest hurdles was handling the different cases of multi-sig setups. I needed to consider all possibilities-for instance, a transaction can be executed directly if only one signer is required or if it’s the final signature needed, while in other cases it can only be proposed or approved. I had to dive deep into the Safe SDK documentation to understand transaction workflows, which differ greatly from traditional single-user transactions. I opened a PR (https://github.com/coinbase/agentkit/pull/330) to the AgentKit repo and am working with the Coinbase team to make these enhancements available in the official release to all onchain agents. As AgentKit is used as plugin for ElizaOS, it will also be available to those agents.
AgentKit is designed as a command-line tool for developers, but I aimed to create an intuitive chat interface for users. To achieve this, I re-architected the project by hosting the AI agent as a backend service and developing a React-based frontend that supports simultaneous multi-user interaction. The frontend communicates in real time with a Node.js backend via socket connections, ensuring prompt processing of user actions. This backend not only initializes and manages the AI agent using AgentKit's action providers but also seamlessly integrates Safe SDK functionalities, wallet providers and blockchain utilities to execute Safe transactions.
Tracks Applied (5)