gmail-dsh

Developers increasingly let DeepSeek Harness agents handle code, documentation, and multi-step workflows, yet a single unread email or a thread that needs a reply still pulls you back to the browser. gmail-dsh plugs Gmail directly into the harness tool runtime: the agent can search the inbox with native Gmail query syntax, read full messages including attachments and thread context, send thread-aware replies, and add or remove labels on individual or entire threads. There is no MCP server and no third-party runtime dependency—the plugin registers straight onto the harness tool, credential, and command systems.

Security is a first-class design constraint. Read operations execute silently; the two mutating tools, gmail_send and gmail_labels, return an approval request from the pre-execute tool hook that the harness approval service turns into a user-facing prompt showing the recipient, subject, or the label change and its target. When no approval service is mounted, mutating operations fail closed to denial, so the plugin is read-only in that deployment and can never send mail without a human who can say yes. OAuth uses Google Desktop client type with a one-shot loopback callback that only shuttles a short-lived authorization code; refresh tokens live solely in the DSH credential store, and multiple accounts can be authorized simultaneously, switched via an optional account parameter.

If your harness workflows already span research, drafting, and follow-up, and you want the agent to naturally check mail, read context, reply on your behalf, or tidy labels without a separate OAuth microservice or desktop app, gmail-dsh is the missing layer. Zero runtime dependencies, in-process credential management, and fail-closed write semantics make it safe in both lightweight local setups and shared team environments.