TL;DR:
Polygon launched the Agent CLI, a command-line tool that unifies in a single installation everything an artificial intelligence agent needs to operate onchain: wallet creation, token sending and swapping, cross-chain bridging, identity registration, and stablecoin payments. The initiative aims to eliminate the infrastructure fragmentation that development teams face when trying to equip their agents with autonomous financial capabilities.
Until now, integrating an agent with the chain required assembling separately a wallet library, a gas abstraction layer, a swap API, a bridge API, and an identity system, each component coming from a different provider with no specific design for the threat models unique to agents. Polygon’s CLI replaces that fragmented stack with an integrated system that installs with a single npm command.
The architecture is built on three layers. The first consists of smart contract wallets with session scope, per-token configurable spending limits, and a 24-hour expiration. Private keys never enter the language model’s context, which neutralizes prompt injection attack vectors aimed at extracting them.
The second is transaction orchestration through Polygon Trails, which handles routing, price discovery, and execution without the agent needing to know which DEX it uses in each operation. The third is native support for ERC-8004, an Ethereum standard for agent identity, co-authored by MetaMask, the Ethereum Foundation, Google, and Coinbase.
The ERC-8004 standard allows each agent to register its identity onchain, accumulate portable reputation, and become discoverable by other agents and services. Complementarily, support for the x402 protocol enables micropayments per HTTP request to be executed directly in stablecoins, with no need to manage API keys or subscriptions. Agents pay exclusively for what they consume in each interaction.
Another important security element is the dry run mode enabled by default: before any transaction is broadcast to the network, Polygon presents a full preview of what will occur. In contexts where an agent can make thousands of decisions, that confirmation step keeps humans in control over what actually touches the chain. The toolkit is compatible with Claude, Openclaw, LangChain, and CrewAI, among other frameworks that support the use of this type of tool.
