Axelar unveiled AgentFlux, an open-source framework designed to run AI agents locally while keeping private keys, trading strategies and client data off the cloud — a pitch aimed squarely at institutions exploring onchain finance but wary of privacy risks.

Developed by Interop Labs, the team behind the Axelar network, AgentFlux lets financial firms deploy “agentic” automation without sending sensitive information to external infrastructure, the company announced Thursday.

The framework tackles one of the biggest frictions in AI-driven crypto operations: tool-calling.

Most agents today rely on cloud-based models to decide which blockchain tools to invoke and how to structure transactions, which can unintentionally expose the very information institutions seek to protect. AgentFlux splits those tasks into two smaller, specialized models — one for picking the right tool, and another for generating the arguments to execute. According to the team behind Axelar, this setup improves tool-calling accuracy by 46% on benchmark tests, bringing local models closer to the performance of larger cloud systems.

Sergey Gorbunov, the co-founder of Axelar, shared with CoinDesk in an interview that he sees AgentFlux benefitting multiple tasks: “First is highly sophisticated trading strategists, they are incredibly proprietary,” he said. “You never definitely want to get, you know, public cloud models to be as good as doing your strategy as you are. And if you upload them, you inherently like violating them. So that's one.”

“The second one I do see is blockchains being used a lot for analysis as well. So for instance, you're doing tax reporting or an investigation of who performed which transactions. Or maybe you have a specific fingerprint and you're expecting malicious activity across a few accounts. Maybe you want to be uploading that fingerprint to a cloud model, or AI system, and asking it to figure out all the correlated transactions, all this suspicious behavior and patterns this account has. So those types of activities you can fully preserve. And run privately,” Gorbunov added.

AgentFlux also plugs into Axelar’s broader multichain strategy. The team positions Axelar as a “gateway to onchain finance,” providing infrastructure for institutions to move assets and data across blockchains from a single integration point. AgentFlux could enable a single AI agent to view risk, assess exposure and transact across multiple ecosystems — a capability the company sees as essential for institutional adoption.

Read more: Canary Capital Files to Launch ETF Tracking Cross-Chain Protocol Axelar