How to Configure the Box MCP Server for AI Agents

Two coding agents pointed at the same repository will happily overwrite each other's work, and neither will notice. The fix is not a smarter agent, it is a shared place to put work where the second one can see what the first one did. In development environments where multiple AI tools like Claude Code, Cursor, and custom LangGraph agents operate simultaneously, file conflicts are common. Developers frequently watch agents undo changes made by a colleague or delete critical files during a code modification run. These coordination failures stem from a lack of a single, shared workspace substrate that manages file state, records transaction history, and exposes files through standard interfaces.

Enterprise teams frequently store their documents and code assets in Box. By exposing these assets to AI agents via the Model Context Protocol, developers can supply their models with real-time content. However, connecting agents directly to storage requires careful configuration. Standard cloud sync systems are designed for human collaboration and do not handle the rapid read-and-write cycles of AI coding assistants. To bridge this gap, teams need a secure connection method that respects corporate permissions while offering a structured data access layer. Establishing this link is the first step toward building a functional agent room where humans and AI tools work side by side.

When developer resources first introduced the Box Model Context Protocol integration, many tutorials assumed developers needed to clone a repository and run a local Node.js or Python server. This local-only setup creates significant operational overhead. Developers must manage local Node processes, configure firewall ports, and handle localhost proxy routing on every team member's workstation. Additionally, a local server cannot easily share access tokens across multiple agent sessions, leading to constant authentication prompts and broken connection states. Running isolated local servers on each developer's laptop also introduces security risks, as access credentials are scattered across multiple local environments rather than managed in a single console.

To address these limitations, Box provides a remote Box MCP server hosted directly at https://mcp.box.com. This hosted architecture represents a significant improvement because it removes the requirement to run a local node server. Instead of running background processes locally, the AI client communicates directly with Box's hosted endpoint. The remote server acts as a secure proxy, executing Box AI operations, searching folders, and downloading files while strictly enforcing the permissions of the authenticated user. This hosted approach simplifies deployment, ensures that security updates are applied automatically, and provides a stable connection for cloud-based agents. Developers do not need to install Node dependencies or debug local connection scripts. They only need to point their client to the hosted URL and complete the OAuth handshake.

Connecting your AI client to the hosted Box MCP server requires an enterprise administrator to authorize the application. Because the server accesses corporate content, Box blocks anonymous connections. The connection relies on standard OAuth 2.0 credentials to authenticate the agent and scope its access.

To authorize the integration, follow this three-step OAuth authorization process:

• Step 1: Admin Console Integration Setup: A Box Enterprise administrator must log in to the Box Admin Console, navigate to the Integrations section, locate the Custom Box MCP Server app, and click Configure. The admin registers the specific Redirect URI provided by the AI client (such as Cursor or a local callback URL) and selects the required scopes.
• Step 2: Credential Generation: Within the integration settings panel, the administrator generates a Client ID and Client Secret. These credentials represent the specific connection between the AI client platform and the enterprise Box tenant.
• Step 3: User Grant and Token Exchange: The developer inputs the Client ID and Client Secret into their AI client settings. When the agent attempts to connect, the client redirects the user to the Box login page at https://account.box.com/api/oauth2/authorize to sign in, approve the requested scopes, and exchange the resulting authorization code for access tokens.

During configuration, the administrator must enable specific access scopes depending on the agent's tasks. The root_readwrite scope allows the agent to search folders, download files, and write new outputs. If you plan to use Box's document intelligence features, you must enable the ai.readwrite scope. This allows the agent to send files to Box AI for summarization and metadata extraction, leveraging Box's internal AI models without downloading the files to the local client. By gating this configuration in the Admin Console, enterprise security teams retain full visibility over which external AI agents can read and write files within the corporate repository.

Coordinate Box MCP Server Workflows in Fastio

Import your files via cloud import, configure workspaces with native endpoints, and manage your Box MCP server workflows. Start your 14-day free trial.

Start 14-Day Trial

Once the administrator generates the credentials, developers can connect their local coding tools to the hosted Box endpoint. The setup process varies depending on whether you are using a command-line assistant or a desktop editor.

For developers using Claude Code, the command-line interface provides a direct utility to configure remote servers. You can run the registration command in your terminal, specifying the HTTP transport option and the hosted URL:

claude mcp add --transport http box https://mcp.box.com

When you run this command, Claude Code registers the remote endpoint. However, because the Box MCP server requires OAuth authentication, the CLI will prompt you to complete the OAuth flow. The tool will open your default browser, directing you to the Box authorization page. After you authenticate and approve the connection, the CLI receives the access token and saves the connection details in the user configuration folder. You can verify the connection status by running claude mcp list in your terminal.

If you are using Cursor, you must configure the connection details in the settings panel. Open Cursor, navigate to the features tab, and find the MCP settings. Add a new server with the following settings:

• Name: box-remote-mcp
• Type: http
• URL: https://mcp.box.com

To pass the required OAuth credentials, you must include your Client ID and Client Secret. Cursor will initiate the browser login flow when the agent first attempts to read a Box file. Once authenticated, Cursor manages the token refresh cycle in the background, allowing the agent to run search queries and edit documents across sessions. The client-side configuration acts as the link between your local editor and the remote Box content, giving your agent direct access to documentation and schemas without manual copying.

While the Box MCP server provides direct access to files, relying solely on Box for active agent coordination highlights clear functional differences. Box is designed for long-term document archiving and human file sync. It does not provide the immediate, interactive workspaces that agents need to collaborate in real time. For example, if Claude Code and Cursor agents are working on the same project, they need neutral ground to coordinate tasks, review outputs, and hand off files to human managers. General storage solutions lack the built-in intelligence layer and multiplayer notes that make human-agent collaboration effective.

This is where Fastio workspaces offer a dedicated coordination layer. Rather than replacing Box, developers can import files from Box using Fastio's cloud import feature, which preserves folder structures during ingestion. Once imported, teams can manage the active phase of their projects within Fastio shared workspaces. This ensures that Box remains the corporate repository of record, while Fastio serves as the active scratchpad and collaboration space where agents and humans share the same file context.

Unlike commodity storage backends, Fastio is built for human-agent collaboration:

• Collaborative Notes: Humans and agents co-edit the same files in real time. Notes feature live multiplayer cursors, allowing an agent to draft a code specification while a developer edits the text in the same window. This eliminates the latency of copying files back and forth.
• Version History: Every file edit is recorded in the version history. If a Cursor agent writes an incorrect script, you can view the changes and restore a prior version. This makes agent operations fully auditable and prevents data loss.
• Granular Permissions: You can restrict agent access at the organization, workspace, folder, or file level, ensuring that sensitive folders remain private. This scopes the agent's context to the specific task at hand.
• Metadata Views: Turn raw documents into a live, queryable database. By defining schema columns in natural language, Fastio's AI designs a typed schema (such as Text, Integer, or Date & Time) and extracts structured data from PDFs, images, or scanned records, making the data queryable via MCP. This allows agents to query contract values or policy numbers directly without processing raw files.

Fastio also handles the final stage of agent workflows through a structured ownership transfer mechanism. A coding agent can register a workspace, import Box files, run search indexing, and build project schemas. Once the setup is complete, the agent generates a claim link to transfer ownership of the organization to a human administrator. This ensures that billing and management controls remain under human supervision.

Organizations run on a paid subscription billed against usage-based credits. Plans are structured to fit different team sizes:

• Starter: $29/mo (includes 1 TB of storage and 300,000 credits).
• Business: $99/mo (includes 20 seats, 10 TB of storage, and 1.2 million credits).
• Growth: $299/mo (includes 50 seats, 50 TB of storage, and 4.5 million credits).

Every new organization starts with a 14-day free trial that requires a credit card. The agent can build the initial project during the trial, and when the handoff is complete, the human owner takes over the subscription. This allows developers to prototype agentic workspaces without upfront overhead. Developers can explore the pricing options and configure their storage for agents to integrate their coding tools with their Fastio workspaces. By separating long-term storage in Box from active coordination in Fastio, teams can build a reliable, secure development pipeline that maximizes agent utility.