Start with the part of your current workflow that is missing or repetitive. The plugin name matters less than the access it needs and the work it removes.
- Understand screenshots and visual interfaces
- Look for a vision bridge or visual toolkit when a text-only model needs OCR, layout, grounding, or pixel comparison. Check whether the plugin sends images to another model or service, what credentials it needs, and whether its output preserves the spatial details your task depends on.
- Work inside an existing browser session
- A browser plugin can let DSH navigate pages, inspect rendered state, and use an authenticated session. Review which browser it connects to, how it scopes actions, whether it can read saved state, and how you will confirm sensitive or irreversible operations before the agent performs them.
- Keep knowledge between sessions
- Memory plugins can store preferences, project notes, conversation summaries, or searchable history. Compare local and hosted storage, indexing method, retention controls, export support, and deletion behavior. A memory plugin should make its data path clear before you give it private project context.
- Connect another model provider
- Provider plugins route DSH requests to a different API or local model server. Check endpoint compatibility, supported model names, streaming and tool-call behavior, required environment variables, rate limits, and cost reporting. Confirm that the provider works with the Harness mode and profile you plan to use.
- Change the Web or terminal interface
- UI plugins can add sidebars, file panels, Git views, terminal controls, themes, or an entirely different TUI. Check whether the change runs in the browser, the host process, or both. That determines which profile receives it and whether a refresh or Harness restart is required.
- Run repeatable or scheduled workflows
- Workflow and scheduling plugins can turn a one-off prompt into a reusable sequence with tools, subagents, or timed triggers. Review the trigger source, permissions, concurrency rules, failure handling, and stored state. For unattended work, make sure the plugin records enough evidence to inspect what ran.