Features and capabilities

A closer look at five areas: integrations, skill authoring, browser automation, file attachments, and media embeds. This is the local Linux assistant, not a hosted cloud product.

Integrations

Third-party services connect via OAuth2 (or a bot token). Each connection is a skill with its own tools. Tokens sit in the credential vault. The model never sees raw tokens.

A unified messaging layer exposes platform-agnostic tools (messaging_send, messaging_read, messaging_search) that delegate to Gmail, Slack, and Telegram adapters. Provider-specific tools (for example gmail_archive, slack_add_reaction) go beyond that generic interface. Telegram uses a bot token, not OAuth.

Connect from Settings or by asking the assistant to run the matching setup skill. You bring the client IDs and secrets. See Channels and Privacy and data.

Dynamic skill authoring

When no existing tool covers a need, the assistant can create, test, and keep a new skill at runtime:

Evaluate: draft a TypeScript snippet and test it in a sandbox via evaluate_typescript_code until it passes.
Persist: write the skill under the hatch workspace skills/ directory via scaffold_managed_skill.
Load: skill_load activates it.
Delete: delete_managed_skill removes the directory and index entry.

Tool	Risk	Description
`evaluate_typescript_code`	High	Run a TypeScript snippet in a sandbox
`scaffold_managed_skill`	High	Write a managed skill with SKILL.md frontmatter
`delete_managed_skill`	High	Remove a managed skill directory and index entry

Authoring tools require explicit approval. Skills can declare child relationships with an includes frontmatter field. Managed skills show up in the desktop Settings UI. Walkthrough: Your first skill.

Browser

Web browsing is the bundled Browser skill. Activate with /browser or let the assistant load it. Operations run as assistant browser commands:

Command	Description
`assistant browser navigate`	Go to a URL
`assistant browser snapshot`	List interactive elements
`assistant browser screenshot`	Visual screenshot
`assistant browser click`	Click an element
`assistant browser type`	Type into an input
`assistant browser press-key`	Press a key
`assistant browser wait-for`	Wait for a condition
`assistant browser extract`	Extract page text
`assistant browser fill-credential`	Fill a stored credential into a field
`assistant browser close`	Close the page

Most browser ops are allowed by default. Navigation with allow_private_network=true is high-risk. Check assistant browser --json status before the first browse. Details: Skills and tools.

Attachments

The assistant can attach files and images to replies:

Desktop: inline in the live event stream; the Linux app shows them
Telegram: gateway delivers photo or document (Telegram's size limit applies)
HTTP: message list returns metadata; a separate attachments URL returns the payload

Sources: an attachment directive in the reply, or files produced by a tool. Cap: 100 MB per attachment.

Inline media embeds

The desktop app can preview images and video URLs (YouTube, Vimeo, Loom).

Video uses ephemeral webview storage. Cookies do not persist between sessions.
Video is click to play. Nothing auto-plays.
Images are lazy-loaded.
Video webviews tear down when scrolled offscreen.

Controlled by ui.mediaEmbeds in the workspace config.json.

Setting	Default	Description
`enabled`	`true`	Global toggle for inline embeds
`videoAllowlistDomains`	`["youtube.com", ...]`	Domains allowed to render video
`enabledSince`	timestamp	Only messages after this date show embeds