A plain-language system overview

Captain Who is a local-first desktop Agent application. “Local-first” means that the application manages projects, conversations, tasks, and most runtime state on the local computer. It does not mean every feature is offline, because models, search, MCP, or image generation may connect to services you choose.

Three layers that are easy to understand

Desktop interface
  Displays conversations, files, approvals, settings, and run state
        ↓
Local Host and Core
  Manage windows, file selection, the Agent Loop, permissions, Tools, and persistence
        ↓
External capabilities you configure
  Model APIs, Tavily search, local MCP Servers, target webpages, and more

Desktop interface

The interface presents actionable state but does not treat page input as execution authorization. It requests work through a constrained interface to the local Host and displays authoritative results returned by the backend.

Local Host and Core

The local layer stores projects, conversations, model settings, Scheduled Automations, Skills, and run traces; assembles model context; and executes the Agent Loop. Operating-system capabilities such as windows, system notifications, file pickers, and the built-in browser are also managed by the local Host.

External capabilities

Model requests are sent to the model API URL you configure. Online search sends queries or target URLs to Tavily. A user MCP Server is a stdio child process started locally, but the server itself may still access the network or other data sources. Review the privacy and security boundaries of each service before using it.

How a task flows through the system

1. You write an objective in a project conversation and choose permissions, a model, and optional Skills.
2. The local Core assembles context from history and Tool definitions.
3. The model returns text or proposes Tool calls.
4. The local system validates each Tool and waits for your approval when required.
5. The Tool executes locally, in the browser, or through an external service and returns a safe result to the model.
6. When the loop finishes, messages, state, and usage are written to local storage and shown in the interface.

Multi-Agent adds a persistent parent–child task tree to this flow. Scheduled Automation starts the same kind of root Agent Run at a scheduled time. Skills and MCP change the methods or Tools available to the current Run.

Where data is stored

The application stores its SQLite database and managed files in the user data directory assigned by the operating system. Project source files remain in the project directory you selected. Removing a project deletes the local conversations, messages, and attachment records associated with that project from Captain Who, but does not modify source files in the project directory.

Model Tokens, the Tavily Key, and image-generation credentials each have a separate credential-storage boundary. Signed macOS releases use the system keychain, while the database retains only opaque references. Do not put credentials in prompts, MCP parameters, project files, or screenshots.

The three most important user judgments

• Sending judgment: Which local content may be shared with the configured model or a third-party service?
• Authorization judgment: Should a particular file, command, browser, or MCP action execute?
• Evidence judgment: Is the final answer supported by actual files, tests, sources, or run history?

If you simply want to begin using the product, return to Tutorials. For current boundaries, read Capability limits.