Air-Gapped AI IDE: Secure AI Coding Without Internet Access
Most AI coding tools require a persistent internet connection. For defense contractors, classified environments, and regulated industries, that is a disqualifying requirement. Fabric is the only AI IDE architected for fully air-gapped operation.
An air-gapped AI IDE enables AI-assisted software development in environments with no internet connectivity. Fabric supports fully offline operation with self-hosted language models, zero external dependencies, and complete data sovereignty. All work — conversation history, context, and configurations — remains within the controlled environment and migrates seamlessly if deployment mode changes.
Why Air-Gapped Development Matters
Defense and Classified Environments
Code written for defense systems, intelligence agencies, and classified programs cannot traverse the public internet under any circumstances. SCIF (Sensitive Compartmented Information Facility) requirements mandate physical network isolation. AI coding tools that require cloud connectivity are categorically excluded from these environments.
Regulatory Compliance
ITAR (International Traffic in Arms Regulations) controls restrict where technical data can be processed. FedRAMP authorization requires specific infrastructure boundaries. GDPR and PIPEDA impose data residency requirements. For organizations subject to multiple jurisdictions, the safest path is ensuring data never leaves controlled infrastructure.
AI Sovereignty Concerns
Organizations in Canada, the EU, and other jurisdictions are increasingly concerned about US government authority to compel American AI companies to provide access to user data. The CLOUD Act and similar legislation create legal mechanisms for this. Architectural sovereignty — where data physically cannot be accessed because it never leaves your environment — is the only reliable protection.
Intellectual Property Protection
Sending proprietary source code to a cloud AI service means trusting that the provider will not retain, train on, or expose your code. Policies promise this, but policies can change. Acquisitions happen. Data breaches occur. Air-gapped operation eliminates the trust dependency entirely — your code never reaches an external system.
The Challenge: Cloud-Dependent AI IDEs
Every major AI coding tool — Cursor, GitHub Copilot, Windsurf, Replit, Bolt — requires internet connectivity to function. Their architecture sends your code context to cloud-hosted AI models, processes the request, and returns the result. Without internet access, these tools provide zero AI capability.
Some tools offer partial workarounds. Cursor's Ghost Mode allows local model inference, but it was not designed for air-gapped deployment and lacks feature parity with the cloud experience. GitHub Copilot has no self-hosted option at any tier. Windsurf requires cloud connectivity for all AI operations.
This creates a significant capability gap. Organizations with the most demanding development challenges — defense, intelligence, critical infrastructure — are the least able to benefit from AI-assisted coding. Fabric was built to close this gap.
How Fabric Enables Air-Gapped AI Coding
Self-Hosted Models
Fabric connects to any model served via vLLM, Ollama, or any OpenAI-compatible API endpoint on your local infrastructure. No external API calls. No internet dependency. The model runs on your hardware, within your network boundary.
Complete Offline Operation
Every Fabric feature — chat, agentic mode, autocomplete, codebase analysis — works without internet access. The IDE is distributed as a standalone application that requires no external services to function. License validation can be configured for offline environments.
Work Portability
All conversation history, context, and configurations are stored locally. If deployment requirements change — from air-gapped to cloud, or vice versa — all work migrates seamlessly. Nothing is locked in a vendor's infrastructure.
Supported Models for Air-Gapped Deployment
Fabric works with any model that exposes an OpenAI-compatible API. These open-weight models have been validated for air-gapped coding workflows.
| Model | Parameters | Min. VRAM | Strengths |
|---|---|---|---|
| GLM-5 | Multiple sizes | Varies | Strong multilingual coding, long context |
| Qwen 3.5 397B (MoE) | 397B (17B active) | 8x A100 80GB or GB200 NVL72 | Near-frontier coding capability, 256K context |
| Qwen 3.5 32B | 32B | 2x A100 40GB | Excellent cost-performance ratio for coding |
| Llama 3.3 70B | 70B | 2x A100 80GB | Strong general-purpose coding, widely deployed |
| Mistral Large | 123B | 4x A100 80GB | Strong reasoning, enterprise-oriented |
| DeepSeek-V3 | 671B (37B active) | GB200 NVL72 or 8x A100 80GB | Excellent code generation, MoE efficiency |
| CodeStral | 22B | 1x A100 40GB | Purpose-built for code, fast inference |
The Performance Trade-Off
Honesty matters here: local open-weight models are currently less capable than frontier cloud models like Claude Opus 4, GPT-4o, or Gemini 2.5 Pro for complex reasoning tasks. The gap is meaningful for multi-file architectural changes, subtle bug detection, and nuanced code review.
However, the gap is narrowing rapidly. Qwen 3.5 397B approaches frontier performance on coding benchmarks. GLM-5 and DeepSeek-V3 deliver strong results on code generation tasks. For many common development workflows — autocomplete, boilerplate generation, test writing, straightforward refactoring — the quality difference is negligible.
Fabric lets you calibrate this dial precisely. Use cloud models when policy allows. Switch to local models when sovereignty demands it. The transition is instantaneous — one configuration change, zero downtime, all context preserved. You are not choosing between AI capability and security. You are choosing where on the spectrum your organization needs to operate, and Fabric supports the full range.
Deployment Architecture
Container-Based Deployment
Fabric's model serving layer deploys as Docker containers. Package the model weights, inference engine (vLLM or Ollama), and configuration into a container image. Transfer to the air-gapped environment via approved media. Deploy on any container runtime — Docker, Podman, or containerd.
Kubernetes Orchestration
For multi-team deployments, Fabric's model serving integrates with Kubernetes for scaling, load balancing, and resource management. Helm charts are available for standardized deployment. GPU scheduling via the NVIDIA Device Plugin ensures efficient hardware utilization across multiple models and teams.
VPC and Private Cloud
For organizations with private cloud infrastructure (AWS GovCloud, Azure Government, private OpenStack), Fabric deploys within your VPC boundary. Network security groups restrict all traffic to internal endpoints. No egress rules required — Fabric makes zero outbound connections in air-gapped mode.
Compliance Framework Support
| Framework | Requirement | How Fabric Addresses It |
|---|---|---|
| GDPR | Data residency within EU | On-premise deployment ensures code and AI interactions never leave your EU infrastructure. No data transfer to US or other jurisdictions. |
| PIPEDA | Canadian data sovereignty | Air-gapped deployment on Canadian infrastructure. Zero cross-border data flow. Architectural sovereignty, not just contractual. |
| ITAR | Technical data access controls | All AI processing occurs within ITAR-controlled environment. No technical data is transmitted externally. Self-hosted models process exclusively local data. |
| FedRAMP | Authorized infrastructure boundaries | Fabric deploys within FedRAMP-authorized infrastructure (AWS GovCloud, Azure Government). No external dependencies beyond the authorized boundary. |
| ISO 27001 | Information security management | Air-gapped deployment satisfies the strictest interpretation of access controls, data classification, and asset management requirements. |
| SOC 2 Type II | Security, availability, confidentiality | Fabric's cloud infrastructure is SOC 2 Type II compliant. Air-gapped deployments inherit the security posture of your controlled environment. |
Frequently Asked Questions
Can Cursor or GitHub Copilot work in air-gapped environments?
No. Cursor, GitHub Copilot, and Windsurf all require internet connectivity to function. Their AI processing happens on cloud servers, which means code must traverse the public internet to reach the AI model. Cursor's Ghost Mode allows local model use but does not support the full feature set and was not designed for air-gapped deployment. For environments where internet access is prohibited, these tools are categorically excluded.
What models can Fabric run in an air-gapped environment?
Fabric supports any model that can be served via a vLLM, Ollama, or OpenAI-compatible API endpoint on local infrastructure. Proven options include GLM-5, Qwen 3.5 (including the 397B MoE variant), Llama 3.3, Mistral Large, DeepSeek-V3, and CodeStral. Model selection depends on your available GPU hardware and the quality-performance trade-off your team is willing to accept.
What hardware is required for on-premise AI model hosting?
Requirements vary by model size. A 70B parameter model requires approximately 140GB of GPU VRAM in FP16 (two A100 80GB GPUs or equivalent). Smaller models (7B-14B) can run on a single consumer GPU. Quantized models (FP8, GPTQ, AWQ) reduce VRAM requirements by 30-50%. For enterprise deployment, we recommend at least 4x A100 GPUs to serve a 70B+ model with acceptable latency for a team of 20-50 developers. For maximum performance with frontier-scale MoE models, the NVIDIA GB200 NVL72 provides 13.5TB of unified HBM3e memory and 1.4 exaflops of FP4 compute — enough to serve multiple 400B+ models simultaneously with room to spare.
How does air-gapped AI coding compare to cloud-based AI coding in quality?
Honestly, local models are currently less capable than frontier cloud models like Claude Opus, GPT-4o, or Gemini 2.5 Pro. The gap is narrowing — open models like Qwen 3.5 and GLM-5 deliver strong performance on coding tasks — but a quality trade-off exists. Fabric lets you calibrate this dial: use cloud models when allowed, switch to local models when sovereignty requires it. The transition is seamless, and all context carries over.
Does Fabric support ITAR and FedRAMP compliance?
Fabric's architecture supports deployment within ITAR-controlled and FedRAMP-authorized environments. In air-gapped mode, no data leaves the controlled environment — there are no cloud API calls, no telemetry, no external dependencies. The compliance posture is determined by your infrastructure configuration rather than vendor policy. Fabric's deployment is designed to operate within existing approved infrastructure boundaries.
Can I start with cloud and migrate to air-gapped later?
Yes. This is a core design principle of Fabric. You can start with cloud-hosted models for maximum capability, then reconfigure to on-premise deployment when requirements change — regulatory shifts, new contract requirements, or organizational policy changes. All conversation history, context, project settings, and workflow configurations migrate with you. No work is lost in the transition.
Deploy AI Coding in Your Secure Environment
Fabric is the only AI IDE built for air-gapped, classified, and regulated environments. Talk to our team about on-premise deployment.