Capability limits

This page prevents capabilities that may exist elsewhere in the industry from being mistaken for features currently supported by Captain Who. Numeric limits are safety caps, not recommendations to use the maximum on every task.

Models and network

Model integrations support generic OpenAI-compatible, generic Anthropic-compatible, and DeepSeek V4 Chat configurations. A compatible endpoint must still implement the corresponding protocol behavior.
Online search depends on a user-configured Tavily API Key. The search Tool cannot be used without a Key.
Models and search are not offline capabilities; data is sent to the corresponding configured service.
Token and cost figures are local statistics or estimates, not the provider's final bill.

Projects, files, and attachments

A single attachment selected in the input box can be up to 8 MiB. When adding a message to an active Agent, one message can contain up to 8 attachments totaling no more than 32 MiB.
Agent attachments support currently registered formats including text, images, PDF, .docx, .pptx, .xlsx, .csv, and .tsv.
Legacy .doc files are parsed through a system capability on macOS only. Legacy .ppt and .xls files are not currently supported.
PDF text is not extracted automatically when an attachment is sent. Activate the PDF Skill to extract or render content on demand. If a scanned PDF has no text layer, inspect it through page rendering and visual capabilities.
File Tool capabilities differ from the Files preview in the right sidebar. The Files preview is read-only and does not support editing, creating, renaming, or deleting.
Files text preview supports up to 1 MiB or 5,000 lines, images up to 12 MiB, and PDFs up to 32 MiB. Office files cannot be previewed directly in the Files panel.
Symlinks may appear in the file tree, but cannot be expanded, previewed, copied as real paths, or revealed in the file manager.

File modification and commands

Structured file writes and patches are subject to workspace, file-version, parent-directory, and symlink checks. Automatic approval does not bypass these checks.
Very large file modifications are constrained by managed-draft and patch-size limits. Have the Agent write in stages or reduce the scope of each change.
Commands are classified by their complete statement and risk. Full access still cannot execute operations that are always denied.
Long-running commands use a managed Command Session. Cancellation does not always prove that an external side effect did not occur.

Skills

Skill sources are Bundled, Installed, and the current project's Workspace. The entry file must be named exactly SKILL.md.
By default, one Run can activate up to 8 Skills, with no more than 512 KiB of activated instruction source in total. Very large directories may be truncated with diagnostics.
Workspace Skills are discovered only in the current project. After creating or modifying one, discover it again in a new Run.
User installation mainly supports GitHub and authorized local directories.
Skill Scripts currently support only scripts/*.py and the Host's Python 3. Missing dependencies are not installed automatically. Preflight and execution require elevated permission conditions, and every actual execution requires approval.

MCP and browser

User MCP Servers support local stdio only. HTTP, Streamable HTTP, OAuth, environment variables, SecretRef, and custom headers are not yet available.
The Agent product path currently uses MCP Tools only. Resources, Prompts, Sampling, Elicitation, Tasks, and MCP Apps are not available.
An MCP call may enter an unknown-outcome state after a timeout, cancellation, or lost connection. Calls with side effects cannot be replayed automatically.
The built-in browser supports manual browsing and managed Agent automation. Webpage requests for camera, microphone, location, notifications, and similar permissions are denied by default.
Browser Sessions retain cookies and site storage across application sessions until the user selects Clear browsing data.
A browser Artifact is a temporary Run artifact, not a permanent file, and has a maximum lifetime of 24 hours. Export it explicitly when it must be retained longer.

Multi-Agent

The collaboration structure is a parent–child tree, not a general-purpose DAG, graphical workflow, or automatic planner.
The user interacts only with the root Agent. Subagent conversations can be observed only in read-only mode.
By default, the maximum tree depth is 8, each tree can contain up to 64 nodes, and at most 4 Agent Turns execute globally at the same time.
One Agent can have at most one active Turn at a time. Templates are isolated by project, and changes affect only future nodes.
More nodes do not provide a linear speed increase and add model-call and context costs.

Scheduled Automation

The scheduler works only while the application and Core Server are running. It is not an operating-system background service and does not guarantee execution to the second.
It supports fixed minute, hour, and day intervals; daily, weekday, and weekly schedules; and custom recurrence by hour, day, week, month, or year. Cron, one-time schedules, and end dates are not supported.
Multiple schedule points missed while offline are merged into one recovery Run rather than replayed individually.
A Task can have at most one unfinished Run at a time. Multiple Scheduled Automations still share global execution capacity with ordinary root Agents and subagents.
The current UI cannot stop an active Run independently. Pause affects future schedules only; delete requests cancellation.
There is no separate approval timeout, run-history retention setting, or cleanup setting. An in-app alert also does not replace a failed native notification.

Office, Artifacts, and image generation

Creating or editing Word documents, spreadsheets, and presentations depends on the corresponding Bundled Skill and managed build process. Foundational Office Tools mainly inspect, validate, and render files.
General managed Artifacts currently support primarily PNG, JPEG, WebP, and PDF. Office documents are still delivered as workspace files.
Image generation currently has one built-in Seedream profile with a default 2K preset and no arbitrary width-and-height protocol.
Image-generation requests are protected by concurrency and queue limits. An unknown result that may already have incurred a charge is not retried automatically.

When you encounter a limit, reduce the task scope, divide the input, or use an explicit export or save step. Do not expand permissions to bypass resource and safety limits.