Glossary

// the station’s words, in plain English

StarNet uses game words for real machinery. Each entry gives the plain meaning first and, where the app uses a different label, that label in a tag.

PEOPLE THE FLOOR BELTS WORK LEARNING AUTONOMY PLUMBING

People and crew

AGENT: One AI worker with a name, a class, a model, a desk, and its own memory. Every sprite on the floor is one. Runs are real model calls with real tools.
OVERSEER: The first agent you create. It runs the station and recruits every agent after it. It can’t be deleted.
COMMANDER: You. The dossier the crew keeps on its commander is what “it studies you” means.
CLASS RECRUITMENT BAY: A head start, not a cage: a Researcher comes tuned for digging, an Engineer for code. 35 curated classes on the roster, a 20-deep SPECIALIST ARCHIVE, plus any custom classes you build or save from an agent.
CREW: All your agents together. Gear you place is shared by the whole crew; only a computer is per agent.
SKIN: The character an agent wears. 36 to pick from, any class, changeable any time in the agent’s CONFIG tab. Cosmetic only.
DELEGATION: One agent dispatching subtasks to crew it summons, then folding the results back. You watch the whole tree run on the floor.

The floor

STATION: Your whole world: rooms, hallways, decks, props, belts, and the crew inside it. Its layout is real configuration, not decoration.
REFIT REFIT STATION: Build mode. Ten tools on keys 0–9: select, room, hallway, surface, move, delete, prop, belt, dupe, lines. Undo always works.
GEAR / PROP: A placed object. Some gear grants a real ability to the whole crew: DISH = WEB, INTEL CAB = FILES, WORKBENCH = TERMINAL, CORE = MEMORY, STUDIO = IMAGES, CONSOLE = COMPUTE. Place one of each and the station is fully powered.
BAY a desk: Where one agent sits. Every bay is another agent working in parallel. An agent’s bay needs a computer in its room, or the floor says NO COMPUTE.
COMMS: The chat panel where you direct the station and where every beat — rate the work, keep this memory, confirm this quest — arrives.
TRUTHFUL TELEMETRY: The station’s core law: it never shows state the harness can’t prove. A lit desk is a live run; the timer is the real one; a dark belt really won’t run.

Belts and lines

BELT: The path work takes across the floor. Laid with the BELT tool (7): click one machine, then another. Lit means “this will run”; dark means something is missing.
LINE: A whole job’s route from front door to out tray: machines plus the belts between them. Every line has a budget ($2.00 per job unless you raise it).
BOX WORK ITEM: One job in transit. When you see a box move, an agent is really working right then. The app’s tooltips call it a work item.
INBOX the front door: Where outside work — a routine, a message from a paired chat — enters and drops onto a belt. Its card holds the line’s name and budget. COMMS orders skip it and land at the agent’s bay.
OUTBOX the out tray: Where finished work leaves the line. Click it to read and rate finished runs.
FILTER a signpost: Reads each job’s text and sends it down a lane: CODE, RESEARCH, or EVERYTHING ELSE. Never drops work. Needs a default lane.
SPLITTER: Fans work across its lanes. Alone: jobs take turns (load-balance). With a JOINER downstream: every lane gets the same job (fan-out).
JOINER: A barrier that waits for every branch (10 minutes for a late one, adjustable), then sends one merged crate on.
MERGER: A lane funnel: several belts in, one belt out, every crate rides straight through. Combines lanes, never jobs. Nothing to configure.
LOOP: The one tile that makes a cycle legal: a DONE lane and a BACK lane, up to 5 passes by default (20 max), optionally until a reviewer says APPROVED.
LANE: One belt leaving a signpost’s tile. A filter or splitter needs at least two; a joiner needs at least two coming in.
LINES / BLUEPRINT LINES tool, key 9: Nineteen ready-made layouts — FRONT DESK, RESEARCH LINE, REVISION LOOP, SORTING OFFICE, PARALLEL CREW, THE GAUNTLET and more — dropped in one click.
ROUTING PLAN: What your belts compile into: the exact plan the server executes. The box you watch and the run that happens are the same plan.

Work

RUN: One agent doing one job: a real model call with real tools, a real timer, and a real cost line in the ledger.
DELIVERY: What a run produced — a file, a report, a reply. Lands in the OUTBOX or in COMMS for you to read and rate.
RATE THE WORK: The beat after any run that did real work: ▲ nailed it mints XP; ◆ close and ▼ missed are noted with no XP and no penalty.
RECIPE WORK › RECIPES: WHAT to run: a ready-made job with a small form. 101 built in; save your own.
ROUTINE WORK › AUTOMATION: WHEN it runs: a job on a schedule. Drops in at the INBOX. Only fires while scheduling is enabled.
SKILL BUILD › ABILITIES: HOW to do it: a procedure an agent consults mid-task when the work matches. 73 built in, more written by agents themselves or installed from a SKILL.md you inspected.
TASKS: Where live work sits — the queue you can see.

Learning

DOSSIER CREW › COMMANDER: What the station has learned about you, in nine dimensions (identity, stack, goals, working style, standing orders, pain points, ambitions, people, schedule). Every agent’s briefing folds it in. Local-first; pin, edit, or forget anything.
BELIEF: One thing the dossier holds, traced to where it came from. Proposed after runs; you Keep, Edit, or Discard.
MEMORY CORE AGENT DOSSIER › MEMORY: An agent’s own long-term notebook. Each memory traces to the run that earned it. Trust grows with use and halves every 30 days unused; low trust is de-weighted, never silently dropped.
XP / LEVEL: Minted only by your nailed it verdict. Levels record what shipped and gate nothing. Level 2 at 50 XP, 5 at 500, 10 at 2,250.
QUEST LOG: Your goals broken into 3–5 milestones, advanced only by real completed work. ◆ YOU ARE HERE marks the open step.
RECOMMENDER: Suggests one thing at a time — a recipe, a class, a bespoke specialist — and only with evidence. A suggestion that can’t say why is dropped.

Autonomy

NIGHT SHIFT: The station choosing a small grounded job while you’re away (30 minutes idle), about every 45 minutes, inside your leash. Off by default.
INITIATIVE: Does it start work on its own? WAIT (default) · SUGGEST · BUILD · FREE.
REACH: How far an unattended action may go. OBSERVE (read only) · SANDBOX (default; nothing leaves the machine) · SEND & PUBLISH.
PACE / LEASH: How many unattended jobs per day: LIGHT 1, STANDARD 3 (default), BUSY 6, MAX 12.
MORNING REPORT: What COMMS prints when you return: what fired, what was skipped and why, drafts to review. Every line maps to a real record.
DECISION LEDGER RECENT DECISIONS: An append-only trail of every autonomous fire, skip, and defer with the gate that decided it.
BUDGET: Hard USD caps — per run, per agent, per day, global — enforced against the real ledger, on top of each line’s own budget.

Plumbing

SIDECAR: The small local server that does the actual work, on 127.0.0.1. The app window talks to it; nothing talks to a StarNet server.
WORKSPACE: Where your station state lives on disk: transcripts, memory, ledgers, routines. On Windows under %LOCALAPPDATA%\StarNet; on macOS under ~/Library/Application Support.
PROVIDER: Who runs the model. Eighteen supported, OpenRouter by default; three sign in with no key; Ollama is free and fully local. STARNET is the optional prepaid one.
BYOK: Bring your own key. Your key, your provider’s bill, real cost shown per run. No markup.
LEDGER: The append-only record of every run’s true spend. What the cost readouts are read from.
CHANNEL: A paired chat app — Telegram, Discord, Slack, Matrix, Signal — where your agents appear as contacts and take orders.
OWNER PAIRING: Telegram only: a one-time 10-minute code you send the bot as /pair, so the first stranger to message it doesn’t own it.
CONNECTOR / MCP: A way to plug in an outside service (Gmail, a calendar, a docs tool) or to let another harness drive your station. Granted explicitly, never assumed.
SERVICE KEYS the KEYS tab: Keys for the services your agents call — kept locally, separate from your model provider key.
LIVE VOICE: Hands-free talking to the station on a speech engine built into the app. No key, no account, your voice never leaves the machine.