Capabilities Guide
Captain Who capabilities are organized in several layers. Distinguishing them helps avoid treating “documented,” “connected,” and “authorized to run” as the same thing.
| Capability | Plain-language meaning | What you primarily do |
|---|---|---|
| Tool | The agent's hands and eyes | Define the goal; inspect tool activity and approvals |
| Skill | Task instructions and supporting resources | Select, install, enable, or update a skill |
| MCP | A general-purpose connection to external tools | Configure a local server and authorize its launch and calls |
| Multi-Agent | A collaborative group managed by a root agent | Define the division of work, observe subagents, and handle approvals |
| Scheduled Automation | Starts a root-agent task on a schedule | Set the schedule, goal, permissions, and notifications |
| Browser Automation | An agent controls managed web pages | Enable the capability, prepare a page, and approve sensitive actions |
| Office and Artifacts | Work with documents, spreadsheets, presentations, PDFs, and images | Choose a skill, provide inputs, and verify the output |
How capabilities work together
A task may use several capability layers. For example, “Summarize industry developments and update a spreadsheet every week” could work as follows:
- A scheduled automation starts a root agent weekly.
- The agent uses a web tool to research sources.
- It activates the Spreadsheets Skill and updates the workbook according to that workflow.
- It delegates research to multiple subagents if appropriate.
- File writes remain subject to the scheduled automation's frozen permissions and approval rules.
These layers do not automatically authorize one another. Installing a skill does not enable MCP. Allowing automatic calls to an MCP server does not expand file permissions. A scheduled automation also does not bypass ordinary agent approvals or safety checks.
If you are unsure which capability to choose
- For a fixed built-in action, start with Tools.
- For a particular method, template, or Office workflow, use a Skill.
- To connect a third-party local tool, use MCP.
- When independent subtasks can run in parallel, use Multi-Agent.
- When the same task should repeat on a schedule, use a Scheduled Automation.
For a more detailed comparison, see Choosing the Right Capability.