Sharing a coding agent brings a small vocabulary with it. Most of the words are ordinary, which is exactly why they get used loosely. This page fixes what each one means in muti, in one paragraph each, so a definition can be quoted without the rest of the post.
Agent room
An agent room is a shared space where a coding agent running on one person’s machine can be reached by a small group of people that person trusts. The room holds the conversation, the request timeline and the approvals. The agent, its subscription, its files and its credentials stay on the owner’s machine. Nothing is shared by default; the room is the boundary, and nothing crosses between rooms. The long version is What is an agent room?
Agent
An agent is a runtime such as Claude Code or Codex pinned to one working directory on one machine, with a handle, an avatar, an online status, a permission mode and an inbox. One machine can host several agents; one agent can sit in several rooms. Agents are named for the job they do (api, site, notes) rather than for the model behind them.
Owner
The owner is the person whose machine an agent runs on. The owner approves requests from others, decides standing grants, and is the only person who can approve one of the four dangerous action types or answer a credential prompt. The owner is not subject to permission modes on their own agent, but their actions are recorded like everyone else’s.
Member
A member is a person in a room who is not its owner. Members send requests, see the timeline and receive results. What a member may run is decided by approvals and standing grants, never by the fact of being in the room.
Inbox
The inbox is the per-agent queue where requests wait while the machine is offline or busy. Each agent processes its inbox in order, one request at a time; different agents on the same machine run in parallel. The owner can remove or reorder queued items.
Request
A request is one task sent to an agent. It has a public id such as R20260906-01 and a status: pending, approved, running, done, failed, cancelled or expired. A request can be cancelled, retried, or resumed under the same id after an interruption.
Timeline
The timeline is the live, persisted record of a request: messages, tool calls with argument summaries, file diffs as paths and line counts, commands with exit codes and output tails, artifacts (files, images, links), approvals, takeovers and errors. It streams to everyone in the room as it happens and stays there afterwards.
Permission mode
A permission mode is the agent’s own rule set for what runs without asking, inherited from the runtime: for Claude Code, plan, default, acceptEdits and bypassPermissions. It is chosen per agent by the owner. See Claude Code permission modes, explained.
Standing grant
A standing grant is an owner’s decision that a specific person may run a specific agent in a specific room without per-request approval. It is revocable at any time; revoking it cancels that person’s queued requests. The four dangerous action types pause even under a standing grant.
Approval
An approval is a pause in a run that waits for the owner to approve, reject or continue read-only, optionally with a note that carries the same weight as the request. Approvals are raised for pushing to a shared branch, deleting, spending money and touching credentials, regardless of trust level, and expire after a set time if unanswered. See four actions an agent should always ask before doing.
Takeover
A takeover is an agent’s request for a human during a run. Credential-related takeovers (logins, two-factor codes, key access) route to the owner. Intent questions (“did you mean staging?”) route to the requester. Both see the exchange in the same timeline.
Borrow an agent
To borrow an agent is to use someone else’s agent, on their machine and their subscription, from your own context, under a grant they can revoke. The borrower gets usage, the timeline and the results, never the account, the credentials or the files. See Borrowing an agent.
Daemon
The daemon is the small background process on the owner’s machine that detects installed runtimes, polls the room service over an outbound connection, runs requests locally and uploads timeline events. It never uploads files, environment variables or keys, and it opens no ports. Installed as a service, it starts at login and restarts after a crash.
Machine
A machine is a computer paired to a muti account by its owner. Nothing runs until the owner approves the machine in the web app. Machines report their local and public addresses so an owner with several can tell them apart, and can be revoked at any time.
Relay
A relay is a server that holds the shared event log in Block's Buzz: humans and agents connect to it over WebSocket, sign what they send with their own keys, and every message, review, workflow step and git event lands in the same log. It is the self-hosted counterpart to a room service: you run it, and you own the history. See Block's Buzz and muti.
AI member
An AI member is the workspace vocabulary for an agent that has been given a name, a role and durable memory so people address it the way they address a colleague. Kylon uses it for agents that are "members, not features"; Buzz gives each one its own keypair and audit trail; muti calls the same thing an agent with a handle, an inbox and a permission mode. In every case the point is that the agent is addressable by more than one person.
Workspace app
A workspace app is a table, database or small internal tool built inside a room, with its own data and version history, so an agent can read and write business records where the conversation happens. Kylon's Apps are the clearest example. muti has none by design: rooms hold messages, timelines, approvals and artifacts, and business data stays where it already lives.
Personal agent
A personal agent is an AI agent that belongs to one person: it keeps memory about them, holds connections to their apps, works on a computer of its own and acts on a schedule or on events without waiting to be asked. Meta Muse, OpenAI dots, Manus Cue and Grok Bot are the 2026 examples; Tencent WorkBuddy, Alibaba's QwenWork and QoderWake add group chats where colleagues can reach it. muti is not one: it gives an agent you already run a room, so other people can send it work. See personal agents in 2026, compared.
Agent computer
An agent computer is the machine an agent works on: a browser, a file system and a terminal it can use the way a person would, so it can reach sites and apps with no API. Muse's Secure VM, a dot's cloud computer and Grok Bot's per-account machine live in the vendor's cloud and need your logins there; WorkBuddy, QoderWake and muti use the computer you already work on, where the logins already are. See allow once, always allow, deny.
Artifact
An artifact is a file, image or link an agent produced during a request. Artifacts appear in the room’s outputs for download, including images the agent generated outside its working directory.
If a term you use is missing, the FAQ on the home page covers the product-level questions, and each post above goes deeper on one concept.
Published · Updated