Settings reference
Open Settings from the local profile entry point in the lower-left corner. The search box on the left filters Settings destinations by page name or group.
Page overview
| Page | Main content | Important notes |
|---|---|---|
| General | Language, permission modes, Custom permissions, context-usage display | Default permissions are always available; Full and Custom modes can be disabled |
| Profile | Local avatar, display name, and ID | Used only for the local interface profile |
| Appearance | Light and dark modes, themes, diff previews, sidebar effects | macOS font smoothing appears only on supported platforms |
| Configuration | Models, providers, online search, image generation | Contains API address and credential-related settings |
| Personalization | Working style, response tone, custom instructions | Custom instructions affect all tasks on this Host and require selecting Save |
| Usage and estimated costs | Tokens, requests, model breakdown, estimated cost | Estimates are not provider bills and do not distinguish currencies |
| Skills | View, enable, disable, install, update, and uninstall Skills | Preview an installation package before confirming it |
| Subagents | Manage subagent templates by project | Template changes affect only subagents created in the future |
| MCP | Manage built-in capabilities and local MCP Servers | Save, startup authorization, enablement, and tool approval are separate actions |
| Environments | Add or remove local projects | Removing a project does not delete the source directory, but does delete associated local application data |
| Archived conversations | Filter, restore, or permanently delete archived conversations | Permanent deletion cannot be undone within the application |
General: permissions
Default
Allows reads and writes inside the workspace. Commands and file modifications require approval at important points and remain under the protected command-safety policy. Suitable for most tasks.
Full access
Expands supported reads and writes outside the workspace, automatically approves commands and patches, and uses a broader command-safety policy. Always-denied operations and path validation still apply. The interface asks for confirmation again when this mode is selected.
Custom
You can configure separately:
- Reads: workspace only or all supported locations.
- Writes: denied, workspace only, or all supported locations.
- File edits: require approval or approve automatically.
- Commands: require approval or approve automatically.
Custom mode still uses the protected upper bound for command safety and is not equivalent to Full access. If writes are denied, automatic approval for file edits is unavailable as well.
For the current conversation, select the task's permission mode in the input box. When an Automation is saved, its resolved permissions are frozen. Later changes to this page do not automatically rewrite existing tasks.
Configuration: models and Providers
The following model integration types are currently supported:
- generic OpenAI-compatible Chat;
- generic Anthropic-compatible Messages;
- DeepSeek V4 Chat.
The global URL and API Token provide a connection for models that still use the default. An individual model can instead use its own paired URL and Token. A model entry can specify a display name, context window, image-input support, and locally estimated price per 1k input, cached-input, and output tokens. Configure the context window to match the actual service; an incorrect value affects capacity judgments.
A model appears in model pickers for conversations, Automation, and subagent templates only after it is enabled. Deleting or disabling a model referenced by another configuration may cause the corresponding task to require repair.
DeepSeek settings also provide reasoning mode and reasoning effort. Actual support still depends on the connected service.
Configuration: online search
Online search requires both enabling Allow and entering a Tavily API Key. Queries and target URLs are sent to Tavily. Signed macOS releases store the Key in the system keychain, while SQLite retains only an opaque reference. The interface shows only a configured, missing, or unavailable status and never displays the stored value.
Configuration: image generation
Image generation currently uses one built-in Seedream profile. You can enter an HTTPS endpoint, model, and API Key, and configure capabilities, the default 2K preset, and a watermark. Configuration complete in the interface does not prove remote connectivity or validate the key. Remote errors become visible during the first actual generation.
Skills
Bundled Skills are provided with the application. Installed Skills can be installed from GitHub or a local directory you authorize. The installation flow parses a candidate, shows its files and compatibility issues, and then requests confirmation. Workspace Skills reside in a specific project and are not installed from the global management page.
Enabling a Skill does not grant file, command, or network permissions. For the exact distinction, see Skill and MCP.
Subagents
Select a project before creating a template. Name and instructions are required, the model must be enabled, and the description helps the main Agent understand the template's role. A template can be disabled, edited, or deleted, but existing subagents continue to use the snapshot taken at creation.
MCP
The Built-in section currently contains browser-automation policy. The External section manages local stdio Servers added by the user. An external-server configuration includes a name, executable, arguments, working directory, and tool-call policy.
Use Ask every time initially. Allow tools to run automatically still does not bypass server-identity, tool-catalog, or risk checks. Deny all calls is an explicit stop switch.
Environments and data deletion
Add project authorizes a workspace through the system directory picker. Removing a project permanently deletes its associated project configuration, conversations, messages, and attachment records from Captain Who, but does not modify source files in the selected directory.
Archived conversations can filter by project and restore conversations. Deleting one or all archived conversations is permanent local deletion. First confirm whether their history and attachments must be retained.
Usage and estimated costs
You can view token trends for the last 7 days, last 30 days, or last year, and filter by model. Only messages whose model has a valid configured price are included in estimated costs. Clearing statistics deletes usage-statistic records only; it does not remove token details already present in historical conversations.