what is it?

netsky went dark during a self-refactor. The key innovation of netsky is giving an always-on daemon access to its source code and a resilient system for self-updates after editing it. This allows netsky to continuously add functionality to itself and, in theory, converge across machines using some interesting methodologies.

Anyway, I’ve gone back and forth on tons of details about how the system should work. I’m a bit frustrated by the myriad differences across LLM tooling. Some examples:

  • with a ChatGPT subscription, you can’t use GPT-5.5’s 1M-context window
  • without a Claude.ai subscription, you can’t use Claude Code channels (Slack, iMessage, etc.)
  • without a ChatGPT subscription, you can’t use the GSuite plugins in the Desktop app
  • …and so on

What’s annoying is how dead-simple and standards-driven all of this shit is. GSuite plugins are hitting standard Google APIs after some OAuth-based authentication.

What makes netsky awesome is it can build that software for itself, at a layer above the agentic CLIs. Don’t depend on Codex CLI to have parity for iMessage support like Claude Code has (as long as you don’t use their API via API key!). Don’t depend on the GSuite having a standard CLI that just works across LLM providers, just have netsky build one for itself and iterate.

But this begs the question: why use their agent harness at all? And the simple answer is economics: I can get a ton more usage off my $200/month plan than if I paid API prices. The cost is dealing with Claude Code having no app server (but Codex does!) and no Rust SDK. Codex doesn’t have async channels (but my fork does! enjoy the 20 minute compile time!). The drudgery continues.

I am somewhat excited for the subsidized AI prices to end so the actual economics can shake out, and I can decide what to build for netsky. I have made some fun architectural decisions:

  • web app via HTML/CSS/JS
    • dashboards via plotly JS + DuckDB WASM
  • desktop app wrapping web app via Tauri
  • meta.db convention for all metadata
  • netsky CLI for all agentic interactions
  • netsky as a control plane
  • iroh for cross-machine netsky constellation communications
  • various data models & useful machinery for building the VSM of netsky

I’m not sure about model (as in LLM) collapse, but perhaps agent harness collapse is on the horizon.