Concepts
The vocabulary of Cyberine in one place, defined once each - fleet, worker, judge, mode, seat, lane, loop worker, route, board, autopilot, draft, inbox and handoff.
Cyberine dispatches one task to several agents, judges what they return, and ships the winner. These are the words its commands, output and skills use. Each is defined once; the commands that touch it are named so you can try it.
Fleet
The set of workers a run may use. cyberine init pins a fleet for a repo in .cyberine/fleet.json, and build.fleet in .cyberine/config.json is the preference for which workers a build run uses. cyberine run --fleet <ids> pins exactly which workers run for one dispatch and is a hard limit. cyberine fleet doctor validates the pinned fleet.
Worker
One agent that receives the task and returns a result. There are two kinds. A loop worker runs a tool-calling loop against a model API. A CLI worker spawns a headless coding agent (claude, codex, muse, Gemini, Command Code and others) as a child process, wrapped in a host sandbox (Seatbelt or bubblewrap) when one is available. Setting CYBERINE_TASK_TRUSTED=false makes the sandbox mandatory.
Loop worker
A worker that Cyberine itself drives: it sends the prompt to one OpenAI-compatible endpoint (a gateway, DeepSeek direct, OpenRouter and similar), runs the tool-calling loop with Cyberine's own tools, and returns a result. It is billed against the provider API key you configured. Its counterpart is the CLI worker, which hands control to a vendor's own binary and runs under that binary's login.
Judge
The component that reads every worker's result and picks a winner. DiffJudge applies each candidate diff in a throwaway git worktree and keeps one that applies and typechecks. ConsensusJudge is a pure majority vote, LlmJudge asks a model, and AdversarialJudge sends the candidates past several skeptics. The result is a verdict, and for a code run the winner's diff is stored in the run record.
Mode
A preset over fleet size, judge chain and verify gate, chosen with cyberine run --mode compete|swarm|deep|smart|rush|harvest or the CYBERINE_MODE variable. compete races the full fleet and judges with the diff and behavior judges. swarm uses one worker per chunk of a plan, then merges. deep uses reasoning-tier models, smart uses flash-tier models, rush runs one worker on the cheapest models, and harvest is the harvest-judge path. With no mode given, the router picks.
Route
The router's decision for a task: its tier (trivial, standard or complex), its kind, the mode, which fleet runs and why. cyberine route "<task>" prints it as a free dry run without dispatching anything.
Seat
One Claude subscription account you have declared on this machine. A seat is not an API key: the work runs by spawning the official claude binary under that account's own login, and Cyberine never reads or stores the credential. cyberine init writes one claude-<seat> entry per declared seat, and cyberine run --fleet claude-<seat> sends a job to that account. Per-account work runs only through the in-process one-shot claude -p worker, so the account that launched a job is the account that pays for it.
Lane
Two related uses. In a fleet, a lane is the entry for one worker; the bare claude worker expands at dispatch time into one claude-<seat> lane per enabled seat. In --mode swarm, a lane is a group of plan chunks that share files or depend on each other, run in order in one worktree while separate lanes run in parallel.
Board
The task list for a repo, a manifest at .local/tasks/ACTIVE.md (or .claude/TASKS.md) that cyberine tasks reads and writes. cyberine tasks shows it, cyberine next lists the runnable tasks, and the writer verbs (add, save, status, note, decide, end) change it. Hand edits are not the supported path: every write goes through cyberine tasks. The format is in the task manifest spec.
Autopilot
cyberine autopilot works the board without you at the keyboard. It takes the runnable tasks, dispatches them, reviews the results and lands winners, until nothing agent-runnable is left or a limit is hit (--max-iter, --max-hours). --dry-run prints the plan and launches nothing. It refuses until the repo records the Cyberine rules as applied.
Draft
A proposed task that exists only in the inbox. cyberine draft "<intent>" writes a tagged brief and then runs propose, critique and revise rounds on a worker to produce reviewed drafts. It spends quota, and it never touches the board.
Inbox
The holding area for drafts. cyberine tasks inbox lists them with the draft report. A draft becomes a task only when you save it with cyberine tasks save, and cyberine tasks inbox reject <draft...> --reason "<why>" discards drafts; a later cyberine draft treats a rejected draft as a duplicate and does not propose it again.
Handoff
A note that lets the next session of the same repo resume where this one stopped. It records Where, Decisions, Open and Next. cyberine handoff write writes a short breadcrumb for free, cyberine handoff pull reads it, and files live under .cyberine/handoffs/. In Claude Code, /cyberine:handoff writes the short note and /cyberine:end wraps the board and writes the full one with one background model turn.
Where to go next
- Getting started: install, a first run, the board.
- Usage: the recipe for every command above.
Beta guide
What the closed Cyberine beta asks you to try, what it uses on your machine, how to upgrade, exactly what to delete to remove it, and where to report a problem.
Cyberine — Supported Features
Cyberine is a vendor-independent, multi-provider AI agent orchestrator (ports-and-adapters): dispatch one task to N agents — raw-API loop workers and/or headless CLI coding agents — judge the...