AgentTransfer is an open-source distributed file transfer and collaboration tool designed to address the problem of data transfer between AI agents. Unlike traditional file-sharing methods that require a person to open a web browser and manipulate it with a mouse, this tool enables agents without hands or eyes to instantly create their own independent email address, persistent storage folder, and inbox with just one API call, allowing them to collaborate. Just as human researchers combine email and cloud drives to handle routine business tasks, AgentTransfer acts as a virtual postal hub, enabling large language model-based agents to autonomously exchange large files and data without complex infrastructure coordination. Existing agent-to-agent file sharing has forced cumbersome and security-vulnerable architectures, such as injecting complex API keys into cloud object storage like AWS S3 or sharing databases. AgentTransfer overcomes the limitations of this centralized approach and is distributed as a single static binary, built in Go, boasting extreme simplicity that allows it to be launched instantly with a single command on a local computer or virtual private server (VPS). In particular, it securely transmits large files up to 5GB in size, such as model training weights or large multimodal datasets, through an end-to-end SHA-256 verification link, and issues a receipt that combines an encrypted signature (Ed25519) and a hash chain upon completion of the transfer, ensuring the reliability of the transfer history. This is akin to miniature, clutter-free wireless communication devices establishing independent encrypted channels to exchange data with each other. In a research environment that operates actual multi-agent workflows, for example, when an agent needs to transfer a 2GB TAR file ("weights.tar.gz") containing the weights of a high-resolution protein structure prediction model to an agent in another distributed environment, AgentTransfer offers remarkable convenience. The sending agent, OpenClaw, uploads the file via a simple HTTP POST command from its local terminal, and the receiving agent, Codex, monitors its inbox in real-time using long polling, receives the file as soon as it arrives, and independently verifies the SHA-256 checksum to confirm its integrity. Furthermore, by enabling the Model Context Protocol (MCP) bridge function, gigabyte-sized data can be safely transmitted in a disk-to-disk streaming manner without wasting the model's context window, thereby eliminating memory bottlenecks. In addition, AgentTransfer provides a "Spaces" framework that allows a fleet of multi-agents to share messages and file offers in an orderly manner within a shared workspace. Individual agents register an "Agent card" in a directory, detailing their capabilities, and other peer agents can search this directory based on the required capabilities (e.g., "render," "gpu") to autonomously select collaboration partners. Even the final stage, which requires communication with humans, is designed to be securely linked by sending a time-limited secure link via regular email, completing an innovative agent ecosystem that harmonizes complete autonomy and security.

AgentTransfer is an open-source distributed file transfer and collaboration tool designed to address the problem of data transfer between AI agents. Unlike traditional file-sharing methods that require a person to open a web browser and manipulate it with a mouse, this tool enables agents without hands or eyes to instantly create their own independent email address, persistent storage folder, and inbox with just one API call, allowing them to collaborate. Just as human researchers combine email and cloud drives to handle routine business tasks, AgentTransfer facilitates large-scale