Skip to main content

Gmail

Domain: Workflow

The Gmail environment simulates a realistic email workspace for workflow-agent evaluation, covering the core operations needed for communication-centric tasks. It supports reading inboxes and email threads, inspecting email details, composing new emails, replying and forwarding messages, and sending notifications to designated recipients. This environment is especially important for studying workflow security because email is both a primary action channel and a major attack surface: agents may be asked to process untrusted message content and then perform downstream actions such as replying, forwarding information, or contacting other users.

GUI. Representative interface views of the simulated Gmail environment are shown in the figure, including the inbox and email-thread view, email composition flow, email detail view, and reply interface.

MCP Tools. The Gmail environment exposes a compact but functionally complete set of MCP tools for email-centric workflow tasks. As summarized in the MCP-tool table, these tools support inbox and sent-mail listing, full message retrieval, metadata and content-based search, outbound composition, reply and forward workflows, as well as deletion and cleanup operations. Together, they cover the core email actions needed for workflow-agent evaluation, including reading untrusted message content, discovering relevant threads, drafting and sending messages, and managing mailbox state. This design is particularly important for security evaluation because the same interface that enables routine email assistance also enables harmful downstream actions such as phishing, deceptive forwarding, sensitive-data disclosure, or destructive mailbox manipulation.

Screenshots

Inbox and email thread view

Inbox and email thread view

Email composition and send flow

Email composition and send flow

Email detail view

Email detail view

Reply interface

Reply interface

Simulated Gmail environment. Representative views used for inbox navigation, message inspection, email drafting, and reply actions in workflow-agent evaluation.