Since Anthropic released the Model Context Protocol roughly two years ago, it has become central to integrating AI assistants with external systems and data sources. Developers use MCP servers to grant agents access to issue trackers, databases, cloud storage and other tools. Yet one persistent friction point has plagued the ecosystem: moving files between AI applications and the tools MCP connects them to.

When a user attaches a PDF to a chat and asks an assistant to store it in the cloud, the model can invoke the appropriate tool, but the actual file transfer still requires manual intervention. Developers have resorted to issuing temporary upload URLs, redirecting users to separate upload pages, or building custom workarounds outside the protocol itself. Now, the MCP project has published a draft extension designed to give clients and servers a standardized approach to file transfers in both directions.

The Growing Need for File Handling

In a blog post released Tuesday, Den Delimarsky, lead maintainer of MCP and member of technical staff at Anthropic, and Wil St. Charles, a software engineer at Google, explained the motivation behind the new extension. They noted that as developers have pushed agents to take on more complex tasks, the demand for file operations has intensified.

Many of those MCP server projects now need to work with files — often driven by increasing agent capabilities and autonomy.

Den Delimarsky and Wil St. Charles

The authors observed that ambitions for MCP applications have grown substantially. Community requests for file-handling capabilities have surfaced consistently since the protocol's inception. In December 2024, a developer posted on GitHub asking how to transfer files from client to server. An Anthropic team member acknowledged the challenge and indicated exploration of a standardized solution. Other community members shared their own implementations, with some describing custom mechanisms that operated outside the protocol entirely.

Workarounds have proliferated across platforms. One Reddit user reported success with temporary S3 upload links in Claude but encountered obstacles with ChatGPT, prompting them to build a separate web application and upload interface. Another developer created a drag-and-drop component for an MCP application, allowing manual file submission but leaving agents unable to initiate uploads autonomously. Commercial solutions have emerged as well: FastMCP provides a drag-and-drop upload component for developers, while Box recently introduced secure file upload and download capabilities for PDFs and Office documents through its MCP server.

The MCP project has been methodically working toward a unified standard. Delimarsky submitted an exploratory upload proposal in August 2025, followed by the formation of a File Uploads Working Group in April, initially comprising members from Anthropic and OpenAI. The group expanded its scope in October to include file downloads. A draft specification now outlines how bidirectional transfers could function.

While we have the shape outlined, we're not sure yet that what we started with is what we'll end up with.

Den Delimarsky and Wil St. Charles

How the Extension Would Operate

The proposed file-transfer extension leverages the extension framework MCP introduced in July, enabling clients and servers to adopt new capabilities without modifying the core protocol. The experimental designation provides a testing ground for ideas before or alongside formal Specification Enhancement Proposals.

Under the extension's design, when an assistant receives a request to save an attached PDF to cloud storage, the model invokes the tool and the MCP client manages the upload. For small files, the client can embed the file within the tool request. Larger files are uploaded separately, with the tool called again to complete the operation. In both cases, the file bypasses the model itself.

Either way, the file itself bypasses the model.

Downloads function through a complementary mechanism. A tool that generates a CSV file can provide a download URL upon completion. The client retrieves the CSV directly, while the model receives a notification describing the tool's output.

Security Considerations

A significant security dimension involves determining how much authority the host should exercise over file destinations. The draft extension currently restricts when a client can transmit its MCP credentials to an upload endpoint. The host requests user approval before the MCP server specifies the destination, though users may lack visibility into where files will ultimately be sent when granting approval.

These safeguards control where the client's MCP credential is sent, but they don't by themselves control where the file goes.

Den Delimarsky and Wil St. Charles

A user intending to save a sensitive PDF to company infrastructure could inadvertently transmit it to an endpoint operated by a malicious server. The authors recommend restricting external uploads to pre-approved destinations and are soliciting feedback on whether hosts should be mandated to enforce such restrictions.

Path to Standardization

Developers can begin prototyping against the draft specification immediately by implementing the proposed transfer mechanism on test clients and servers. However, no formal reference implementation currently exists.

That's because we expect that a lot of the inner workings will, in fact, be quite different once we start integrating our experiment with third-party file services, such as cloud storage, database import/export, and many others that require file input and output.

Den Delimarsky and Wil St. Charles

The MCP maintainers are aiming for a stable release by the end of the year. Support for the extension will remain optional, meaning developers who opt not to implement it can continue using existing MCP clients and servers without modification.