starnet

Open source alternative to Claude Managed Agents

1.5K stars
MIT
Active this week
Last commit 6 days ago

Last verified Oct 20, 2026

Run a local AI agent crew from a pixel-art station on your desktop. Bring your own API keys; agent state, transcripts, and spend logs stay on your machine.

Visit website GitHub1.5K

Read the full breakdownWho it is for, how it works, trade-offs, the quick start and the FAQ

who it's for

Who starnet is for#

Developers automating multi-step local workflows

StarNet gives you a ready-built platform with scheduling, tool consent, budget controls, and MCP connector support without writing the infrastructure yourself. Wire agents to your existing tools via MCP, put recurring tasks on cron schedules, and retrieve completed work from the OUTBOX as files.

Skip if:

Skip if you need a web-based or team-shared agent workspace; StarNet is a desktop application with no multi-user or cloud-sync path in the current release.

Power users who run agents away from the desk

Night Shift lets you leave the station running and come back to completed work. Agents operate inside an adjustable leash, every away-action is logged, and Task Brief questions arrive over Telegram, Discord, or another connected channel so you can unblock an agent from your phone.

Skip if:

Skip if your tasks require constant approval at every step; Night Shift is designed for bounded autonomous runs, not continuous hand-holding.

Privacy-focused users who cannot use cloud agent services

For work involving confidential documents, client data, or internal tooling, keeping transcripts and instructions off third-party servers is a hard requirement. StarNet routes model calls directly from your machine and stores nothing externally.

Skip if:

Skip if your team needs shared agent libraries, centralized logging, or a managed service; StarNet is a per-machine local runtime with no built-in multi-user sync.

Teams migrating from OpenClaw or Hermes

StarNet can import an existing agent from an on-disk OpenClaw or Hermes workspace, minting a StarNet agent from the persona, instructions, memory, and model it finds. API keys do not transfer and must be re-entered in the KEYS tab.

Skip if:

Skip if you are migrating from a cloud agent platform with no local workspace format; StarNet's import covers only OpenClaw and Hermes on-disk formats.

the problem

The problem it solves#

Running AI agents at useful scale means either writing your own agent framework from scratch or handing your workflow to a managed service. The first option is a significant engineering project: wiring model providers, building a persistence layer, managing tool consent, handling scheduling, and writing budget controls. Most people never finish it.

The managed services make the setup instant but introduce a different problem: your API calls route through their servers, your transcripts live in their database, and your agent instructions are stored outside your control. For any workflow involving private documents, client data, or internal tooling, that tradeoff is unacceptable. You also pay per run or per seat on top of the API costs you are already paying to your model provider.

how starnet solves it

How it solves it#

Pixel-art station as live runtime view

The station renders the actual state of the runtime: each room is a capability-scoped team, each hallway is an authorized handoff lane, and a placed object is a real capability grant. The layout is not a dashboard of your workflow; it is the workflow itself. Moving or removing a room changes what agents can do.

Multi-agent crews with concurrent runs

Each agent in the station gets its own workspace, transcript, memory, and bounded permissions. Start with one agent and place bays or summon specialists to add more, each running concurrently as a genuinely distinct agent with its own model, persona, and tool loadout.

Bring your own key, any provider

Paste an OpenRouter key to access hundreds of models, or sign in with Anthropic, OpenAI, or Google. Keys live in the OS keychain, never in the frontend. For zero-cost runs, connect Ollama on 127.0.0.1:11434 and StarNet reports it ready once it can list your local models.

Night Shift with explicit activity leash

Leave the station running and agents keep working within an adjustable activity budget. Every action taken while you are away is logged and reviewable. If ambiguity arises, Task Briefs surface a concrete question over your connected channel before the agent proceeds.

Connected messaging channels

Wire your station to Telegram, Discord, Slack, Signal, or Matrix. Talk to agents away from the desk and receive Task Brief questions over whichever channel you are already watching. Completed deliverables land in the OUTBOX as real files, not buried in chat history.

MCP connectors and cron scheduling

Attach Model Context Protocol servers and OAuth connectors to extend what agents can reach. Launch multi-step recipes, grant reusable skills, and put recurring work on cron schedules with visible output. A spend ledger and run history persist on disk and are shown as-is.

strengths · trade-offs

Strengths and trade-offs#

Strengths

All data stays on your machineAgent memory, transcripts, spend records, tasks, and schedules persist on disk under the local StarNet workspace. Provider requests leave your machine only when an agent makes a model call; no intermediate cloud service touches the payload or stores a copy. This makes StarNet viable for sensitive documents and internal tooling that cannot leave your infrastructure.
MIT license with no usage restrictionsYou can fork, modify, and redistribute the code commercially. The MIT license has no usage restrictions or service terms to navigate. The StarNet name, logo, and artwork are not included, so forks need their own branding, but the code itself carries no licensing barrier.
Free local model support via OllamaInstall Ollama, pull a model such as llama3.1, and point StarNet at it for zero API cost. StarNet talks to Ollama on 127.0.0.1:11434 and only marks it ready once it can list your local models. This is a genuine path to running agents with no API costs for development and lower-stakes tasks.
Secrets stay in the OS keychainAPI keys are held by the sidecar and the OS keychain, never in the frontend or persisted in a config file the UI reads. The architecture separates the local authority (sidecar, keychain) from the rendering layer, so a compromised frontend cannot read your keys.

Trade-offs

-Windows and macOS only; no Linux public releaseThe public release train supports Windows 10/11 and macOS. Linux packages exist as internal build artifacts but are not a supported public release target. If your team uses Linux desktops as primary machines, StarNet is not a viable option in the current release.
-macOS has less real-world coverage than WindowsThe README explicitly notes that Windows is the most-tested desktop target and macOS has less real-world coverage. Both Mac builds pass Developer ID checks and Apple notarization, but early adopters on macOS should expect more rough edges than on Windows.
-BYOK means API costs are the user's responsibilityStarNet does not bundle any model API access. You need an OpenRouter key or direct provider credentials, and every model call is billed to your account. The spend ledger is on disk and transparent, but teams with variable agent workloads need to set budgets carefully to avoid unexpected API bills.

versus alternatives

starnet vs alternatives#

StarNet vs Cloud AI Agent Platforms

Most commercial AI agent tools in this space are managed services: you build agents in their interface, they run on their infrastructure, and your data (transcripts, instructions, API call logs) lives in their database. StarNet takes a different deployment path as a local desktop runtime.

StarNet vs Microsoft Copilot Studio

Microsoft Copilot Studio is a low-code AI agent builder inside the Microsoft 365 ecosystem. It handles infrastructure, authentication, and deployment for you and integrates natively with Teams, SharePoint, and other Microsoft services.

Feature	StarNet	Microsoft Copilot Studio
License	MIT	Proprietary
Runs locally	Yes	No (cloud only)
Data storage	On your machine	Microsoft cloud
Model choice	Any (BYOK, Ollama)	Primarily Azure OpenAI
Pricing	Free (your API costs)	Per-user licensing ($30+/mo)

StarNet is the better choice when you need local data storage, model flexibility beyond Azure OpenAI, or you work outside the Microsoft 365 ecosystem. Copilot Studio is worth choosing if your team is already in Microsoft 365 and you need out-of-the-box Teams integration with no infrastructure to manage.

StarNet vs Relevance AI

Relevance AI is a no-code AI agent builder with a visual workflow editor, multi-step task automation, and team collaboration features delivered as a managed SaaS service.

Feature	StarNet	Relevance AI
License	MIT	Proprietary
Runs locally	Yes	No (cloud only)
Data storage	On your machine	Relevance AI cloud
Model choice	Any (BYOK, Ollama)	Supported providers only
Pricing	Free (your API costs)	$19/mo+ per user

StarNet is stronger for users who need full data ownership, want to run local models via Ollama at zero API cost, or need a tool they can extend or modify. Relevance AI is a better fit when you need team collaboration features, shared agent libraries, and a polished no-code interface without managing any local installation.

install · quick start

Quick start#

bash

Running from source requires Node.js 18+ and Git; the sidecar starts with no additional install steps.

```bash
git clone https://github.com/androoAGI/starnet.git
cd starnet
node sidecar/index.js
```

tech stack · detected from GitHub

What it's built on#

Languages: JavaScriptRust

frequently asked

FAQ#

Is StarNet free to use?

StarNet is free to download and run; the code is MIT licensed with no usage fees or seat licensing. You pay only for the API calls made by your agents, billed directly by whichever model provider you connect (OpenRouter, Anthropic, OpenAI, Google). Use Ollama as the provider for zero API cost.

Does StarNet work with local models?

Yes. Install Ollama, pull a model such as llama3.1, and select OLLAMA as the provider in the first-run brain screen or in Settings. StarNet talks to Ollama on 127.0.0.1:11434 and only reports it ready once it can list your local models. Local models are smaller than cloud ones, so expect slower results on long tasks.

What platforms does StarNet run on?

StarNet supports Windows 10/11 (64-bit) and macOS (both Apple Silicon and Intel). Linux packages exist as internal build artifacts but are not a supported public release target. Windows is the most-tested target; macOS has less real-world coverage as of the current release.

How does StarNet keep my data private?

All agent state, transcripts, memory, spend records, and schedules persist on disk in the local StarNet workspace on your machine. Provider requests leave your machine only when an agent makes a model call; no intermediate cloud service stores a copy. API keys are held by the sidecar and the OS keychain, never in the frontend.

Can StarNet run multiple agents at the same time?

Yes. You can place multiple agents in the station, each with its own workspace, transcript, memory, and bounded permissions, and run them concurrently. Each is a genuinely distinct agent run with separate model calls, not a simulation. The visual station projects live runtime state so you can see what each agent is doing.

also worth a look