Features & Capabilities - Developer Guide

A deep dive into the platform's major feature areas — integrations, skill authoring, browser automation, file attachments, and media embeds.

Integrations

Vellum integrates with third-party services via OAuth2, each exposed as a bundled skill with its own set of tools.

The unified messaging layer provides platform-agnostic tools (messaging_send, messaging_read, messaging_search) that delegate to provider adapters for Gmail, Slack, and Telegram. Platform-specific tools (e.g. gmail_archive, slack_add_reaction) extend beyond the generic interface.

Connect via the Settings UI or the integration_connect HTTP endpoint. OAuth2 tokens are stored in the credential vault — the LLM never sees raw tokens. Telegram uses a bot token (not OAuth).

Dynamic Skill Authoring

The assistant can create, test, and persist new skills at runtime when no existing tool covers a user's need.

Evaluate — drafts a TypeScript snippet, tests in a sandbox via evaluate_typescript_code. Iterates until it passes.
Persist — writes the skill to ~/.vellum/workspace/skills/<id>/ via scaffold_managed_skill.
Load — calls skill_load to activate the new skill.
Delete — removes via delete_managed_skill.

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

All authoring tools require explicit user approval. Skills can declare child relationships via the includes frontmatter field. Managed skills appear in the macOS Settings UI.

Browser

Web browsing is provided by the bundled browser skill. Activate via /browser or let the agent load it automatically.

Tool	Description
`browser_navigate`	Navigate to a URL
`browser_snapshot`	List interactive elements
`browser_screenshot`	Take a visual screenshot
`browser_click`	Click an element
`browser_type`	Type text into an input
`browser_press_key`	Press a keyboard key
`browser_wait_for`	Wait for a condition
`browser_extract`	Extract page text content
`browser_fill_credential`	Fill a stored credential into a form field
`browser_close`	Close the browser page

All browser tools are auto-allowed by default. browser_navigate with allow_private_network=true is elevated to high-risk.

Attachments

The assistant attaches files and images to replies, delivered across three channels:

Desktop — inline base64 in SSE events; macOS app renders thumbnails
Telegram — gateway delivers via sendPhoto/sendDocument (20 MB limit)
HTTP API — GET /v1/assistants/:id/messages returns metadata; GET /v1/assistants/:assistantId/attachments/:attachmentId returns full payload

Sources: <vellum-attachment> directives in response text, or auto-converted from tool output. Limit: 100 MB per attachment.

Inline Media Embeds

The desktop app renders inline previews for images and video URLs (YouTube, Vimeo, Loom).

Videos use ephemeral webview storage — no cookies persist between sessions
Videos require click to play; nothing auto-plays
Images are lazy-loaded
Video webviews are torn down when scrolled offscreen

Controlled via ui.mediaEmbeds in ~/.vellum/workspace/config.json.

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