User Guide

This documentation is for people using Captain Who for the first time. You do not need to understand large language models or programming terminology before you begin. Complete a task first, then learn as needed why an agent can read files, call tools, and continue through multiple steps.

Captain Who is a local-first desktop AI work assistant: the app, project permissions, conversation records, and most runtime state are managed on your device. Model requests, web searches, and third-party services you choose to connect may still send necessary data to those providers. “Local-first” does not mean “fully offline.”

Recommended learning paths

Get started in 10 minutes

1. Read the Product Overview to decide whether Captain Who fits your work.
2. Follow Installation and First Launch to obtain the version currently available to you.
3. Use Complete Your First Task to configure a model and start a conversation.
4. If you want to work with a folder or code repository, continue to Use Your First Project.

Browse by feature

• Everyday Use: Conversations, files, Git, online access, models, permissions, and history.
• Capabilities Guide: Tools, skills, MCP, multi-agent collaboration, scheduled automations, browser automation, and Office artifacts.
• Task Tutorials: Complete end-to-end workflows built around real goals.
• Best Practices: Describe tasks more clearly, manage large projects, and use agents more safely.
• Reference: Quickly look up settings, statuses, keyboard shortcuts, capability boundaries, and terminology.
• FAQ: Start with a symptom and find the relevant resolution.

Move from using agents to understanding them

Start with From Chatbot to Agent, then read The Agent Loop, Context Management, and Principles of Multi-Agent Collaboration. These articles focus on intuition and examples; they do not require you to read code.

Keep three things in mind

1. What an agent can do depends on the current model, project, permissions, skills, and MCP configuration. A capability appearing in the interface does not mean the current task has permission to use it.
2. Check the target and impact before approving an action. Pay particular attention to commands, file changes, sensitive browser actions, skill installation, and calls to external MCP servers.
3. External actions cannot be guaranteed to roll back. Stopping a task can prevent later steps, but files already written, requests already sent, and third-party actions already triggered may remain in effect.

Documentation boundaries

These docs cover behavior that users can observe and control. They do not disclose internal protocols, database structures, or maintenance procedures. Developer documentation and public user documentation are maintained separately. If the interface differs from these docs, first confirm the app version, then rely on the capabilities actually shown as available in that version.