Glossary
Every fleetd term, A to Z. Each one links to the page that uses it most.
Agent key
Section titled “Agent key”A key that lets an MCP client or script call the hub, with a role (read, operate or admin). It starts with
fld_ and is shown once, or written to a file only you can read. What it writes in a chat shows as that agent’s,
never yours. See MCP clients.
Agents strip
Section titled “Agents strip”The compact line under the lead’s message in a chat: how many agents are working, one line each. It expands to their lanes and time.
Approval
Section titled “Approval”An item in the chat that shows a gated command with Approve and Deny. See Secrets & approvals.
The plan a lead writes for a piece of work before it starts: what to build, and what done looks like.
The one ongoing conversation with a project’s lead. You give work and answer the lead there. See Chat with the lead.
Critical path
Section titled “Critical path”The chain of work that decides when a piece of work finishes. Shortening it makes it finish sooner. See Where time went.
Decision
Section titled “Decision”Your answer to a lead’s question, logged by the lead with decide in a searchable log. A later question that a
decision already answers is never asked again.
Done-when
Section titled “Done-when”A line in a brief saying what done looks like. Writers and reviewers read it; the lead closes the work, with why, once it holds.
Gated command
Section titled “Gated command”A command a lead wants to run that is irreversible, costs money or uses a credential. It runs only after you approve it in the chat. See Secrets & approvals.
The coordinator of a fleet. It runs on one always-on machine and stores all state. See Concepts.
Human key
Section titled “Human key”A key that belongs to a person’s device, such as the desktop app. Only a human key writes to a lead as you, approves commands, enters secrets or steers: what it sends counts as yours. Every human key belongs to a person.
Invite
Section titled “Invite”A pairing code bound to a new person, or to another device of someone on the hub, from
fleetd invite. It works once, for 7 days. See People & invites.
The long-running Claude session that plans a project’s work and runs its agents. Each project has one, and one chat with you. It never writes code.
Max agents
Section titled “Max agents”The cap on how many agents run at once, per project and across the fleet. Agents over it queue. See Max agents.
Merge train
Section titled “Merge train”The queue that merges a repo’s pull requests after CI passes on the exact commit, in batches when several are ready. Only the train merges. See Merge train.
Needs-you dot
Section titled “Needs-you dot”The dot on a project in the sidebar: the lead’s last message waits for your reply, or a secret or approval item is open.
The fleetd process on a worker. It connects out to the hub and starts the Claude sessions.
Pairing code
Section titled “Pairing code”A one-time code, such as K4QW-7RNP, that connects the desktop app or a new worker to the hub. It lasts 15 minutes.
The app’s code grants the role it was printed with (fleetd pair --role, default admin).
Person
Section titled “Person”Someone on the hub, with a name and a role (owner, admin or member), who owns their devices. What they do is recorded under their name. See People & invites.
Policy file
Section titled “Policy file”An optional .fleetd/policy.yaml in a project’s checkout. It sets tool rules and the branch prefix. See
Config & files.
Project
Section titled “Project”A git repo on a worker that fleetd works on, with one lead and one chat. See Projects.
Receipt
Section titled “Receipt”The small mark under a message you sent while the lead was working: whether it was queued, taken in, or acknowledged. See Steering.
Reviewer
Section titled “Reviewer”A Claude session that reviews a pull request. Reviewers don’t read the PR description; they judge the change.
Screen
Section titled “Screen”A browser, Android emulator, iOS simulator or Linux desktop an agent opens with session_start. You can watch it live,
take it over and replay its recording. See Screens.
Secret
Section titled “Secret”An item in the chat with a masked field. The value is sealed to one machine and never enters the transcript. See Secrets & approvals.
Session
Section titled “Session”One claude process doing one job, such as a lead, a writer or a reviewer.
Room for one running agent. The hub caps how many agents run at once, per project and fleet-wide.
A message to an agent that is already working, such as “skip the docs page”. In a chat, typing while the lead works steers it. See Steering.
One job a lead hands to a session: write code, review a PR, or research.
A message that restarts a lead’s work: your chat message, a task result, a merged PR.
Worker
Section titled “Worker”A machine that runs agents: any machine with a node. See Machines.
Workstream
Section titled “Workstream”fleetd’s record of one piece of the lead’s work, with an id such as W41. The lead keeps them; you never file one.
Each is open until the lead closes it as done, or it is stopped.
Writer
Section titled “Writer”A Claude session that changes code. Each writer gets its own git worktree and branch, and opens a pull request.