Dropbox Reclaim Rebuilds Calendar Assistant with Custom AI Agent Platform

Dropbox-owned Reclaim has rebuilt its calendar assistant to support natural-language scheduling requests without discarding its existing scheduling logic. The new architecture introduces a custom agent platform that connects large language models to Reclaim’s internal tools, allowing users to describe goals in their own words rather than selecting specific settings.

Custom Agent Architecture

To handle open-ended requests like "make time for this tomorrow," Reclaim engineers built a bespoke agent loop instead of relying on broad third-party agent frameworks. This custom system allows the model to interpret user intent while the agent manages context, tools, and instructions. Key components include:

• Contextual Tool Access: The agent receives only relevant calendar details and tools for each request, limiting scope and focusing on immediate tasks.
• Subagents: Complex requests can be delegated to specialized subagents, with internal checklists ensuring the overall goal stays on track.
• Direct Provider Connections: Reclaim connects directly to model providers, enabling faster adoption of new API capabilities and avoiding the lag or excessive structure of generic frameworks.

The platform also supports the Model Context Protocol (MCP) in two ways: an internal MCP client allows Reclaim’s agent to use external tools, while an external MCP server exposes Reclaim’s tools to AI clients like Claude and ChatGPT.

Consistency and Safety via Schedule Actions

A core challenge was ensuring AI-driven changes behaved identically to manual edits and automated scheduling. Reclaim solved this by routing all changes through Schedule Actions, a standard internal operation for tasks like creating events or updating RSVPs. Whether a change originates from a user, the background scheduler, or an AI agent, it uses the same validation and commit process. This prevents fragmentation and ensures that side effects, such as adjusting flexible events or notifying attendees, are handled consistently across all interfaces.

Preview Mode for Reviewable Changes

To prevent unintended calendar disruptions, Reclaim introduced Preview Mode. This feature generates a temporary, non-destructive version of the user’s calendar where AI suggestions can be reviewed before being applied.

• Pure Function Scheduler: The automated scheduler was refactored into a pure function, allowing it to calculate proposed schedules without saving to the live calendar.
• Redis Caching: To maintain speed during repeated preview calculations, Reclaim stores accessible copies of calendar data in Redis, updating attendee availability dynamically as events are moved in the preview.

This approach ensures users can verify how a change affects their entire schedule and shared events before confirming the update.