Monitor AI agent workflows in a pixel-art space station desktop environment
You can see characters moving on screen to understand what tasks the AI is performing, which tools it is using, and the associated costs at a glance. StarNet is a local-first desktop application that places AI agents in the rooms and corridors of a pixel-art space station, visually representing their execution status and permission scopes. Users connect their own model keys and can transparently manage the process of actual model calls and tool executions.
The space station layout is the workflow
The StarNet interface is not mere decoration but a direct projection of the running runtime state. Station rooms represent teams with restricted permissions, corridors represent approved task handoff paths, and placed objects represent actual functional permissions. The moment a user draws the station layout, the workflow for the agents to execute is determined. Starting with a single agent, placing a Bay, or summoning an expert allows each agent to perform tasks simultaneously within their own independent workspace and permission scope. During this process, model calls, tool usage, and cost generation are all recorded as actual data, not simulations.
Model connection and external channel integration
You can connect models by pasting an OpenRouter API key or logging in with Anthropic, OpenAI, or Google accounts. Keys are stored in the OS keychain and are not exposed to the frontend. You can connect agents to various messengers such as Telegram, Discord, Slack, Signal, and Matrix to issue work instructions from outside the desktop. Enabling the 'Night Shift' feature keeps the station in an active state, allowing agents to continue working within specified scopes, with all remote tasks logged for later review. It also supports verified multi-step recipe execution, reusable skill assignment, and cron-schedule-based task scheduling.
Transparent records and output management
All expenditures, budgets, and execution histories are permanently stored on disk and displayed as is. When an agent completes a task, the output is saved as an actual file in the OUTBOX, not in the chat log, allowing immediate viewing. If there are ambiguous parts during a task, 'Task Briefs' generate specific questions and options and send them to the connected channel, preventing incorrect guesses. You can also extend the scope accessible to agents by attaching MCP (Model Context Protocol) servers and OAuth connectors. Voice input is supported via push-to-talk, and output is delivered in the station's consistent voice.
Local execution and security verification
Node.js 18 or higher and Git are required, and Rust and Tauri prerequisites are needed for desktop shell development. The sidecar uses only Node core modules and runs without additional installation; you can start the local server with the node sidecar/index.js command. Installing Ollama to use local models allows free execution without API keys or accounts, but it may be slower than cloud models and quality may drop in long tasks. Public releases support Windows and macOS; Windows installers must pass Authenticode and timestamp verification, and macOS builds must pass Developer ID and Apple notarization to be distributed. Linux packages are internal build artifacts and are not publicly supported.
Pre-use checks and license scope
StarNet releases its code under the MIT license, but brand assets such as the name, logo, and sprite art are owned by Andrew Sims and are not included in the license. Forks or derivative works must use unique names and logos and cannot be disguised as StarNet or labeled as officially endorsed. Currently, Windows is the most tested target, while macOS has relatively less real-world usage coverage. Security vulnerabilities should be reported via the private channel specified in SECURITY.md, not through public issues. When using local models, be aware of performance limitations and confirm that station state and records remain in the local workspace unless network tools or connectors are explicitly used.