A goal becomes real work.
Understand the controls before granting an agent access.
Three execution modes
| Mode | What it does |
|---|---|
| Autonomous agent | A model selects from its granted tools, receives results, delegates bounded subtasks, and returns files or a final response. Requires a funded model key. |
| Team workflow | A saved sequence of specialists. Each stage has its own transcript and can use its permitted tools; outputs are handed to the next stage. |
| Tool-only website research | A deterministic reader and report builder for up to three public URLs. No model inference is needed. Not a general autonomous agent. |
What the widgets mean
The station, workflow view, execution feed, approvals, and file vault share the same mission state. Drag the station to orbit it; select a station or agent-strip button to inspect the specialist. Pause decorative motion without pausing the runtime. Press Ctrl/⌘ K to search workspace actions.
Ambient light, the core globe, and robot idle motion are decorative. Tool events, active status, files, token counts, and credits come from runtime records. The guided sample is separated and explicitly fictional. “Configured” is a configuration flag, not a successful live-service test.
Useful tools
Agents can read public pages, search with an optional Brave key, calculate arithmetic, create/read/list mission files, search approved memory, propose memory updates, delegate to another specialist, and ask the user a question. Owner-only remote MCP calls and local Docker Python are available after separate configuration. No API key enters the model’s prompt.
Files are virtual UTF-8 artifacts scoped to a mission; they are not arbitrary host filesystem access. Python returns bounded stdout/stderr, not arbitrary host files. Open model-generated code as text first; it is not automatically trusted.
Your controls
Agents: create up to 12 specialists and choose capabilities. Workflows: build up to 20 saved sequences of one to six stages, using drag-and-drop or move buttons. Knowledge: upload small text, Markdown, CSV, or JSON sources and attach selected material. Memory: keep user-approved reference notes for later tasks; this does not change model weights.
Files: inspect, download, or remove saved artifacts. Missions: select a task, inspect its events and outputs, export a ZIP, or request a stop. Schedules: set bounded owner jobs only when a worker is configured. Credits: inspect actual ledger entries and unused-budget releases.
Approvals and budgets
Approve or deny the exact tool call shown. A reply can explain your decision but does not change the displayed arguments. Mission-scoped file writes can be preauthorized for that run; memory changes, remote connector calls and local Python still require approval.
The runtime reserves a maximum budget and records model usage. Pilot metering is 1 credit per 1,000 input tokens and 4 credits per 1,000 output tokens, with a 1-credit minimum per model turn. It is not a USD price quote. Provider invoices may differ by model and tool. The operator bears any gap, including auxiliary-service bills. A conservative estimate can stop a task early.
An interrupted model request or approved external connector call can have an unknown outcome. It is held for owner review rather than automatically repeated. Wait at least 120 seconds, check provider/connector records, and record verified additional usage in Settings → Review workspace requests.
Limits of this release
Three concurrent model-driven missions per workspace, 32 model turns per run, bounded delegation depth and tool actions, 24 retained agent runs, 80 small text files, and a workspace size ceiling. Export and archive regularly. Separate missions have isolated files; delegates within a mission share its files and remaining budget.
This is an independently implemented alpha, not full StarNet feature parity. It does not include a native signed installer, computer-use browser automation, direct host shell access, messaging or voice applications, a marketplace, model training, a cryptocurrency checkout, or a guaranteed-correct AI system.
Visual motion controls
Choose Full, Eco or Pause at the bottom of the page. This changes ambient effects only; it does not pause an agent mission. Reduced-motion system preferences and your previous Pause selection are respected. Select Full to explicitly enable the kinetic display. The instrument deck's radar, waveform and mesh are visual effects; its FPS, viewport and clock are actual display readouts, not AI performance claims.