Developers have been hacking around MCP’s file-upload problem for two years. A fix is taking shape.

For the past two years, developers have been devising creative ways to bypass a limitation in the Model Context Protocol (MCP), a tool that enables AI assistants to interact with external tools and data. This issue arises when trying to pass files from an AI app to an MCP tool, such as saving a PDF to cloud storage. While various workarounds have been devised, including temporary upload links or sending users to separate upload pages, MCP developers have outlined a solution to streamline the process.

In a recent blog post, MCP lead maintainer Den Deli marsky and Google software engineer Wil St. Charles explain the need for a standardized method of handling file uploads and downloads. As AI agents become more capable and autonomous, the demand for file handling features has grown. The MCP project aims to address this need by developing a common standard for file transfers.

The proposed extension, currently in the experimental stage, allows clients and servers to opt into the new feature without altering the core protocol. This approach enables developers to prototype and refine the implementation before a formal Specification Enhancement Proposal (SEP). Under the extension, files bypass the model during transfer. When an assistant is asked to save an attached PDF to cloud storage, for example, the model calls the tool, and the MCP client handles the upload.

For larger files, the client uploads them separately and calls the tool again to complete the job. Similarly, when a tool exports a CSV, for instance, it can include a download URL after the file has been generated. The client can then fetch the CSV directly, and the model receives a message informing it of the tool's output.

One significant security consideration involves the control the host has over where files are sent. The proposed extension limits when MCP credentials can be sent to an upload destination. The host asks the user to approve the upload, which means the user may not know where the file is being directed. As the MCP project continues to work on this extension, they aim to find the best solution while continually testing and refining the approach with real-world feedback.