Glossary
Quick definitions. No fluff. Click through to the linked pages for more depth.
| Term | Definition |
|---|---|
| Assistant | Your personal AI entity. Has a name, personality, memory, and the ability to take actions on your behalf. Defined by workspace files and powered by Claude. |
| BOOTSTRAP.md | The onboarding script your assistant follows for your first conversation. Guides it through discovering its name, personality, and your needs. Deleted automatically when onboarding completes. |
| CES | Credential Execution Service. A separate process that handles credential-bearing operations in isolation, so your assistant never sees credential values in plaintext. |
| Channel | A way to communicate with your assistant. The six channels are: vellum (desktop app), telegram, slack, email, and phone. See Channels. |
| Computer use | The ability to see and control your Mac via accessibility APIs and screenshots. Perceive → verify → execute → observe loop. macOS only. |
| Context compaction | Automatic summarization of older messages when a conversation approaches the context window limit. Preserves goals, decisions, and key artifacts while removing filler. |
| Conversation | A persistent message thread with your assistant. Workspace files are loaded at the start. History is stored in SQLite and can be compacted (summarized) when it gets long. |
| Daemon | The background process (TypeScript/Bun) that runs your assistant. Handles conversations, memory, tool execution, and skill management. Communicates with the desktop app via HTTP and SSE. |
| Gateway | The public ingress point that receives webhooks from external channels (Telegram, Slack) and forwards them to the daemon. Keeps the daemon off the public internet. |
| Guardian | You — the primary owner of the assistant. Guardians have full access to memories, workspace files, tools, and credentials. |
| Heartbeat | A periodic background checklist. Your assistant runs through HEARTBEAT.md at a configurable interval and surfaces things that need your attention. See Scheduling. |
| Host tools | Tools that run on your actual machine outside the sandbox (e.g., host_file_read, host_bash). Always require your permission. |
| IDENTITY.md | The file containing your assistant's name, emoji, personality summary, role, and home. Parsed fields are displayed in the app. Loaded into every conversation. See Identity. |
| Interface | A more specific identifier than channel. Interfaces include the channels plus native client apps: macos and cli. A message on the vellum channel might come from the macOS app or the CLI — the interface tells the system which one. See Channels. |
| Invite code | A 6-digit code used to onboard new contacts on external channels (Telegram, Slack, email). The contact redeems the code to verify their identity with your assistant. |
| Journal | Retrospective entries stored as individual .md files in the workspace. The most recent entries are loaded into every conversation for continuity. Older entries are carried forward when they leave context. See Identity. |
| Memory | Information your assistant extracts and saves from conversations for future reference. Seven kinds: identity, preference, constraint, project, decision, event, and journal — each with different lifetimes. Retrieved via hybrid search (semantic + keyword). See Memory & Context. |
| NOW.md (Scratchpad) | A single workspace file your assistant overwrites with its current state: what it's working on, what it's tracking, upcoming items. Loaded into every conversation so the assistant always has context on what's active right now. See Identity. |
| Notification | A message your assistant sends to you proactively — not in response to something you said. Delivered through the notification pipeline to your connected channels. |
| Permission prompt | The dialog that appears when your assistant needs approval. Shows what it wants to do, the risk level, and offers options from one-time allow to persistent rules. On external channels, this becomes an inline button or text-based approval. See The Permissions Model. |
| Platform Managed | A hosting option where your assistant runs in Vellum’s cloud instead of on your machine. Same capabilities and security model, hosted for you. |
| Playbook | Trigger-action automation rules for handling incoming messages. When a message matches a trigger pattern, the assistant automatically responds, drafts, or notifies based on the configured autonomy level. See Scheduling. |
| Private conversation | A conversation with an isolated memory scope. Memories from a private conversation can’t leak into other conversations, but it can still read from your shared memory pool. |
| Risk level | Every tool is classified as low, medium, or high risk. Low-risk tools run automatically; medium and high-risk tools prompt for approval. See The Permissions Model. |
| Sandbox | The security boundary around ~/.vellum/workspace/. Tools running inside the sandbox can read and write workspace files freely. Enforced at the OS level (sandbox-exec on macOS, bubblewrap on Linux). |
| Schedule | A recurring or one-time automated action using cron syntax, RRULE, or ISO 8601 timestamps. Can execute tasks (background conversation) or send notifications. See Scheduling. |
| Skill | A bundle of tools, instructions, and configuration that gives your assistant a specific capability (e.g., Gmail, Browser, Computer Use). 28 are built-in; community skills can be installed from the registry. See Skills & Tools. |
| SOUL.md | The file defining your assistant's personality, core principles, behavioral rules, and preferences. Loaded into every conversation. Editable by you or your assistant. See Identity. |
| Subagent | An autonomous background agent spawned by your assistant for parallel work. Runs independently with its own context, reports back when done. |
| Surface | An interactive visual element (card, table, form, app) that your assistant creates and renders in the chat. Used via the ui_show tool. |
| Task | A work item in your personal task queue. Can be ad-hoc or based on a reusable template. Has priority tiers (high, medium, low) and status tracking. |
| Tool | A single atomic action your assistant can take (e.g., file_read, browser_navigate, gmail_draft). Each tool has a risk level that determines whether it needs your approval. |
| Trust rules | Persistent allow/deny rules stored in ~/.vellum/protected/trust.json. Accumulated from your approval decisions over time. Each rule has a tool, pattern, scope, decision, and priority. |
| Trusted contact | A verified person who can message your assistant through external channels. They can chat but can’t access your memories, modify your workspace, or use sensitive tools without guardian approval. |
| USER.md | The file where your assistant stores structured information about you — name, pronouns, locale, work role, preferences. Loaded into every conversation. See Identity. |
| Watcher | A polling system that monitors external services (Gmail, Google Calendar, GitHub, Linear) for changes and processes them with an action prompt. See Scheduling. |
| Workspace | The folder at ~/.vellum/workspace/ containing your assistant's identity, personality, skills, databases, and configuration. See The Workspace. |
That's the conceptual foundation. Next up: Trust & Security for the full picture on privacy and permissions.
On this page