Introducing AgentFlux, a Privacy-Preserving AI Framework for Onchain Finance

New open-source framework keeps keys, strategies and client data on-device, paving the way for institutional onchain strategies powered by AI

Axelar, the Gateway to Onchain Finance, today announced the release of AgentFlux, a new open-source framework for privacy-preserving, on-device agentic AI. Built to run powerful AI agents locally on consumer-grade or enterprise hardware, AgentFlux is designed to help institutions bring agentic automation into their onchain financial stacks without sending private keys, trading strategies or sensitive client data to the cloud.

Incubated at Interop Labs, the initial developer of the Axelar Network, in collaboration with researchers behind the AgentFlux research paper, the project is being launched as a free software package. A reference trading demo, built on Ethereum using a Coinbase tool-calling environment, showcases an end-to-end, AI-driven onchain trading workflow. 

The research paper is co-authored by Axelar co-founder Sergey Gorbunov and collaborators Rohan Kadekodi, Zhan Jin, Keisuke Kamahori, Yile Gu and Baris Kasikci of University of Washington, Sean Khatiri and Noah H. Bayindirli of Interop labs.

Agents, blockchains and the privacy gap

As institutions experiment with AI-driven trading, portfolio operations and reporting, agents are rapidly becoming a new layer in the financial technology stack. These systems interpret natural-language intent, break it into multi-step tasks and execute those tasks by calling external tools, including wallets, exchanges and onchain applications.

In blockchain ecosystems, that creates a fundamental tension. Historically, users and institutions have enjoyed strong guarantees that keys and transaction construction happen locally: wallets live on user-controlled devices, transactions are built and signed locally, then submitted to the network. With cloud-based agents in the loop, cryptographic material, prompts and contextual data would be sent to remote infrastructure, undermining both the cryptographic guarantees and the privacy properties that drew many participants to blockchains in the first place.

“Interacting with blockchains is inherently complex, and agents are here to abstract that complexity away,” said Sergey Gorbunov, co-founder of Axelar and co-author of the AgentFlux paper. “But if agents require you to ship keys, trading strategies and client context to the cloud, you’ve broken the core promise of self-custody and privacy that blockchains were built on. AgentFlux is about restoring those guarantees in an agent-native world: the agent runs where your keys live.”

A new architecture for on-device agentic AI

AgentFlux introduces a framework in which the most sensitive parts of agentic workloads—reasoning, tool orchestration and context handling—run fully on-device, while heavier, non-sensitive workloads can be selectively delegated to external services when needed.

Technically, the system is built around a decoupled fine-tuning and inference approach for large language model (LLM) tool-calling. Rather than relying on a single monolithic model to decide both what tool to call and how to call it, AgentFlux disaggregates each tool call into two subtasks:

• Tool selection: choosing the correct function or API (for example, get_wallet_details, buy_asset or convert_to_usdc).
  
  
• Argument generation: producing precise, structured arguments for that tool (such as which wallet, which asset and what size of trade).

How AgentFlux works

AgentFlux uses lightweight LoRA adapters to fine-tune specialized models for each subtask, then coordinates them through a dedicated inference framework that dynamically loads the appropriate adapter at each step of a multi-tool workflow. This architecture allows local models to approach cloud-level performance in complex tool-calling scenarios while preserving on-device privacy and efficiency. In experiments on the MCP-Bench benchmark, the AgentFlux methodology improved tool-calling accuracy of a Qwen-2.5-7B base model by 46%, outperforming other local reasoning and fine-tuned models of similar or even larger sizes.

“AgentFlux shows that you don’t need to ship every token to a frontier cloud model just to orchestrate tools correctly,” said Rohan Kadekodi, lead author of the AgentFlux paper. “By splitting the problem into classification and argument generation, and giving each its own optimized adapter, we can run agents locally that behave much more like top-tier cloud systems, without exposing sensitive data.”

Demonstrating an onchain trading workflow

To illustrate how this architecture applies to onchain finance, the initial AgentFlux release includes a trading demo built around a Coinbase tool-calling environment on the Ethereum blockchain.

In this demo, Coinbase exposes a set of tools that an LLM can call, such as “get wallet details,” “check Ethereum balance,” “buy ETH” or “convert to USDC.” A user begins with a natural-language request, like asking an agent to rebalance into SOL or execute a specific strategy. The agent must then:

1. Select the correct sequence of tools (for example: check balance → place trade → verify updated balance).
   
   
2. Generate the right arguments at each step (which wallet or key reference to use, which asset pair, what notional size).

If the agent picks the wrong tool name or passes malformed arguments, the transaction flow fails—a common outcome for generic local models that have not been trained for precise tool-calling. AgentFlux’s specialized adapters, by contrast, are trained separately to make each of these decisions, dramatically increasing the reliability of the end-to-end trading workflow while keeping private keys and context local to the institution’s environment.

The result is a concrete example of an agent-driven onchain trading stack in which inference, planning and transaction orchestration happen under the institution’s control. This architecture can extend to portfolio analytics, treasury management, compliance reporting and other sensitive financial operations.

Built for institutional DeFi and Axelar’s multichain rails

For global institutions exploring onchain finance, privacy, data residency and key custody are not optional. Trading strategies, client portfolios and compliance workflows often span both traditional financial systems and multiple blockchains; any AI system that touches them must respect strict internal controls and regulatory expectations. 

AgentFlux is designed with these constraints in mind. Sensitive workloads and cryptographic material remain on-device or within institution-controlled environments, while the framework allows carefully scoped interactions with external tools and models when necessary. Only anonymized or aggregate outputs need to leave the local environment.

By pairing AgentFlux with Axelar, institutions can extend these private, agent-driven workflows across systems. Axelar is a decentralized network and development platform that securely connects the world’s blockchains and financial infrastructure, providing a single integration point for tokenization, multichain settlement and liquidity orchestration. Branded as the Gateway to Onchain Finance, Axelar enables investors to tokenize, trade and earn yield across dozens of ecosystems through one secure, compliant entry point. 

“Institutions don’t want a different agent for every chain or every venue; they want a single, governed agent that can see their exposure, reason about risk and act across the entire onchain universe,” Gorbunov said. “Axelar provides the multichain rails; AgentFlux provides the privacy-preserving brain that can sit inside an institution’s perimeter. Together, they form a blueprint for how onchain finance will actually operate at scale.” 

Open-source, extensible and ready for builders

AgentFlux is being released as free software under an open-source license, with full code, documentation and evaluation pipelines available publicly. The repository includes:

• A finetuning pipeline for creating classifier and tool-specific adapters.
  
  
• An inference server that exposes a familiar chat-completion interface for applications.
  
  
• An orchestration and evaluation framework to benchmark tool-calling performance across different tool ecosystems.

From the outset, the project is aimed at builders: cryptonative developers, quantitative teams, infrastructure providers and institutional innovation groups. Axelar and Interop Labs are encouraging developers to fork the code, plug in their own tools (from DeFi protocols and market venues to custodial systems and internal data sources) and experiment with new categories of privacy-preserving onchain agents.

“We see AgentFlux as a starting point, not a finished product,” said Gorbunov. “Our goal is to give the community a solid, research-backed framework and a live demo that proves what’s possible. From there, we want developers and institutions to take it in directions we haven’t even imagined yet.”

Accessing AgentFlux

Developers and institutions can get started with AgentFlux today:

• Explore the AgentFlux landing page for an overview of the architecture, demo and motivating use cases.
  
  
• Read the full AgentFlux research paper for technical details on decoupled fine-tuning, LoRA adapters and benchmark results.
  
  
• Visit the AgentFlux GitHub repository to review the code, run the trading demo and fork the framework for custom deployments.

About AgentFlux

AgentFlux is an open-source framework for privacy-preserving, on-device agentic AI. Developed by researchers including Axelar co-founder Sergey Gorbunov, AgentFlux introduces a decoupled fine-tuning and inference methodology that separates tool selection from argument generation, using specialized LoRA adapters and a dedicated inference framework to deliver cloud-comparable tool-calling performance on local hardware. The project is designed as a foundation for trust-preserving agentic applications across finance, Web3, developer tooling and beyond.

About Axelar

Axelar is the Gateway to Onchain Finance, a decentralized network and development platform securely connecting the world's blockchains and financial institutions, developed by Interop Labs and other contributors. Its configurable technology stack enables seamless tokenization, yield and trading products critical for the next wave of institutional-grade use cases.

Alongside partners including top-tier institutions such as Deutsche Bank, Mastercard and Kinexys by J.P. Morgan, Axelar is developing the next generation of financial infrastructure: from global stablecoins to tokenized funds, built with security, compliance and enterprise scale in mind.

For more information, visit axelar.network and the AgentFlux project page.

‍