Code That Cannot Phone Home - Air-Gapped AI Agents
Code That Cannot Phone Home
Some environments have no internet connection by design. Military systems, secure research labs, financial trading floors, medical devices. These air-gapped systems cannot use cloud AI APIs. But they still need automation.
The Air-Gap Constraint
An air-gapped system means no outbound network connections. No API calls to OpenAI. No sending screenshots to a cloud vision model. Everything the AI agent does must happen on the local machine with locally available resources.
This eliminates most current AI agent architectures, which depend on cloud LLMs for reasoning and decision-making. But it does not eliminate the need for automation. People working in air-gapped environments still have repetitive tasks, still need to navigate complex applications, and still benefit from intelligent assistance.
Local Screen Understanding
The accessibility API does not require an internet connection. AXUIElement on macOS provides the complete UI tree using only local system APIs. An agent can understand what is on screen, identify interactive elements, and perform actions - all without any network call.
For reasoning and decision-making, on-device models running through frameworks like MLX or llama.cpp fill the gap. A 7B parameter model running on Apple Silicon handles task classification, instruction following, and simple planning without external dependencies.
The Trade-Offs
Local models are less capable than cloud models. They handle well-defined tasks reliably but struggle with ambiguous instructions or complex multi-step reasoning. The solution is constraining the agent's scope - narrow, well-defined automations rather than general-purpose assistance.
In practice, this matches what air-gapped environments need anyway. The tasks are specific and repeatable: data entry across applications, report generation from multiple sources, system monitoring and alerting. These do not require cutting-edge reasoning - they require reliable execution.
Air-gapped AI agents are not a compromise. They are a different design point that prioritizes reliability and security over capability breadth.
Fazm is an open source macOS AI agent. Open source on GitHub.