Some AI coding agents can run in an air-gapped environment, but only when every required component operates inside the isolated boundary. A local editor or CLI does not make a product air-gapped if inference, sign-in, licensing, telemetry, updates, or tool calls still require the public internet.
Verify the complete dependency chain:
- Inference: The approved model endpoint is hosted and reachable inside the boundary.
- Runtime and identity: Installation, licensing, authentication, and configuration work offline.
- Tools: Repositories, issue systems, documentation, shells, and build services have internal endpoints.
- Dependencies: Registries, base images, compilers, and updates are mirrored or imported through an approved process.
- Operations: Logs, backups, policy enforcement, and incident review remain inside the environment.
Ask the vendor for a current architecture and data-flow diagram for the exact edition you plan to deploy. Confirm which features degrade or fail when outbound traffic is denied. SaaS-only products generally do not meet a strict air-gap requirement unless the vendor offers a separately documented self-hosted or on-premises deployment.
Run a read-only proof in the real restricted network and capture DNS, authentication, inference, tool, update, and telemetry behavior. If one required service cannot be hosted, mirrored, or disabled internally, the answer for that deployment is no. Treat “offline capable,” “local client,” and “air-gapped” as different claims.
Related reading: Is my code safe with AI coding tools? and Does BYOK improve coding-tool privacy?.
