Glossary
Agent
The execution entity that lets a model observe an environment, call Tools, and continue from results in a loop. It does not refer only to a particular model.
Agent Loop
The cycle in which a model chooses the next step, a Tool executes, and the model observes the result and continues until completion, failure, cancellation, or pending approval.
Approval
A user's approval or rejection of one specific frozen action. An approval rationale or model-generated text does not grant additional permissions.
Artifact
An image, PDF, browser download, or other result reference generated by a Tool and managed by the application. Different Artifacts have different retention and export rules.
Attention
A persistent marker in Scheduled Automation that requires user review, such as pending approval, failure, an important update, or invalid configuration. Marking it as resolved does not change the task state.
Scheduled Automation Run
One execution instance of a Scheduled Automation Task, triggered by a scheduled time, a manual action, or recovery.
Scheduled Automation Task
A persistent scheduled-task configuration containing a prompt, target, model, permission snapshot, schedule, and notification policy.
Browser Automation
The capability that lets an Agent navigate, click, read, download, and perform other operations through the application's managed browser. It is different from Scheduled Automation.
Checkpoint
Safe execution state saved at an approval or recovery boundary so the system can continue after revalidation without asking the model to recreate the action.
Compaction
Replacement of the model's view of complete early history with a shorter summary to free context capacity. It does not delete chat records from the interface.
Context Window
The amount of information a model request can process, including system rules, Tool definitions, history, attachment descriptions, Tool results, and reserved output.
Conversation
A chat container that persistently stores user messages, Agent responses, and related run records. The root Agent and every subagent each have their own conversation.
Core Server
Captain Who's locally running core service, responsible for the Agent Loop, permissions, Tool orchestration, scheduled execution, and SQLite data. Users normally do not operate it directly.
Destination
The location where Scheduled Automation results go: a new root chat for every run, or an additional turn in one existing root chat.
Exact Archive
An exact local archive retained for supported execution content and used to review history. Not every external MCP result, browser result, or very large raw result enters the Exact Archive.
Host
The trusted application boundary that actually connects to the operating system, local Core, and managed external capabilities. The model proposes an action; the Host validates and executes it.
MCP
Model Context Protocol, a protocol that lets an application discover and call external tools. Users can currently add local stdio MCP Servers.
Model / Provider
A Model is a specific model identifier. A Provider/Profile describes the protocol and connection configuration used to communicate with a model service. Captain Who can store different configurations for multiple models.
Multi-Agent
The mechanism through which a root Agent delegates independent tasks to subagents in a parent–child tree and coordinates results through persistent messages and follow-up work.
Outcome Unknown
A state in which an action may have produced an external side effect but the system did not receive a confirmable terminal result. Inspect the real environment first; do not retry it as an ordinary failure.
Project
A local workspace authorized by the user for Captain Who, together with its in-application metadata. Removing a project does not delete the source directory, but does delete associated local application data.
Root Agent
The entry point of a Multi-Agent task tree and the Agent with which the user communicates directly and handles approvals.
Run
One Agent execution with a defined start, state, and terminal outcome. A conversation can contain several Runs over time.
Skill
A versioned instruction package with a known source that may contain references, templates, resources, and optional scripts. A Skill teaches an Agent how to work but does not grant permissions.
stdio
A way for local processes to exchange protocol messages through standard input and standard output. User-configured MCP Servers currently use this connection type.
Subagent
An Agent node created by a parent Agent and given a focused task. Users can observe its conversation in read-only mode, but coordinate it and handle approvals in the root conversation.
Token
A unit used by a model to measure text and structured content. It is not equivalent to a character. Context capacity and usage statistics are normally expressed in tokens.
Tool
A structured capability registered for the model by the Host, such as file reading, search, commands, MCP, or collaboration. The model can only propose a call; it cannot grant itself execution permission.
Turn
A logical conversational turn triggered by one input and closed by one result. The Agent Run inside one Turn may contain several model requests and Tool calls.
Workspace Skill
A Skill located under .agents/skills/ in a project and discovered only in that project. It is managed separately from global Installed Skills.