Air-Gapped AI Code Assistant
An AI code assistant provides code completion, generation, review, and refactoring to developers — and in enterprise form, does it without sending proprietary source code to an external model vendor, operating on a network with no connection to the public internet — models, updates, and telemetry all move by controlled offline transfer, so the system functions fully inside a classified or isolated enclave.
Classified software programs write and maintain enormous codebases with zero access to Copilot-class tools — a growing productivity gap against unclassified peers. An air-gapped code assistant closes it with local code models and in-enclave repo retrieval, delivered through the same offline bundle discipline as any other enclave software.
Air-gapped code assistance sits at an awkward intersection: developers are the least tolerant users of latency and the most capable of routing around a tool they dislike, and an enclave is exactly where they cannot route around it. That is an advantage if the assistant is good and a persistent morale problem if it is not, because the comparison they make is against tools they used on their last unclassified project.
The technical consequence is that completion latency dominates every other consideration. A code assistant that answers in 900 milliseconds is a novelty; one that answers in 200 is infrastructure. Inside an enclave you cannot solve this with a bigger hosted model, so the design is a small, fast, code-tuned model resident on local GPUs for completion, with a larger model reserved for chat-style explanation and review where a second of latency is acceptable.
The compensating advantage is repository context. Vendor assistants index your code generically and cautiously; an enclave deployment can index the whole monorepo, its build system, and its internal libraries without an exfiltration review, because none of it leaves. Well-tuned repo retrieval against your own conventions frequently outperforms a stronger model with weaker context.
Why teams run their AI code assistant air-gapped
Built for defense, intelligence, critical-infrastructure and classified-environment teams.
Zero external connectivity, by design
An air-gapped AI code assistant makes no outbound calls — no license pings, no telemetry, no model API fallbacks. If a component phones home, it fails certification; the architecture must assume the internet does not exist.
Built for classified and SCIF environments
Defense, intelligence, and critical-infrastructure operators need AI capability where cloud AI is categorically prohibited. The AI code assistant runs entirely on enclave hardware and clears accreditation reviews because there is nothing external to assess.
Controlled update path
Models, embeddings, and software updates arrive as signed offline bundles through your cross-domain transfer process — the same discipline you already apply to any software entering the enclave.
Core capabilities of an enterprise AI code assistant
Completion & generation
In-IDE completion and chat-based generation served by code-tuned open-weight models on your infrastructure.
Repo-aware context
Retrieval over your codebase gives suggestions that match your architecture and conventions — without indexing code externally.
PR review agents
Agents review pull requests for defects, style, and security patterns before human review.
Policy-safe by construction
Source never leaves the perimeter, satisfying IP counsel and customers whose code you hold under NDA.
What an air-gapped deployment changes
Everything ships as a self-contained bundle: container images, model weights, embedding models, and documentation must install from local media with no registry or CDN access.
Local models only: the AI code assistant serves open-weight models on enclave GPUs; there is no cloud fallback tier, so model selection and routing happen entirely inside the gap.
Audit evidence must be exportable on your terms — logs stay in the enclave and leave only through your controlled review process.
The air-gapped AI code assistant stack
Air-gapped code assistant, layer by layer — with the reason each choice holds up under this deployment mode.
| Layer | Typical choice | Why, here |
|---|---|---|
| Completion model | A small code-tuned open-weight model, quantized | Chosen for latency, not leaderboard position. Fill-in-the-middle support matters more than general reasoning ability for inline completion. |
| Chat & review model | A larger model on the same GPU pool | Explanation, refactoring, and PR review tolerate latency that inline completion does not. Splitting the two is what makes both usable. |
| Repo retrieval | Local index over the monorepo, refreshed on merge | The enclave advantage: you can index everything. Staleness is the risk, so tie re-indexing to the merge pipeline rather than a nightly job. |
| IDE integration | Editor extensions distributed through the enclave's own channel | Marketplace installs do not work offline. Plan extension packaging and updates as part of the bundle, not as a developer self-service step. |
| Build & CI hooks | Review agents triggered from the in-enclave CI server | Enclave CI is usually a self-managed GitLab or Jenkins, so integrations written for cloud CI providers need rework. |
Sizing an air-gapped AI code assistant
| Profile | Scale | Hardware | What actually binds |
|---|---|---|---|
| Small team | 10–30 developers | One GPU hosting the completion model, chat sharing the card | Completion requests are frequent and tiny; they interleave badly with long chat generations on a single card. |
| Engineering org | 100–300 developers | Dedicated completion GPU plus separate chat/review capacity | Separating the two workloads is what keeps p95 completion latency flat during the working day. |
| Multi-repo program | Several codebases, differing access rules | Shared model tier, per-repo indexes with access scoping | Index isolation follows repository permissions — the model tier can be shared, the retrieval scope cannot. |
Regulations that point to air-gapped
Classified handling
The AI code assistant operates inside SCIF and enclave boundaries; there is nothing to accredit outside them.
ITAR / export control
Technical data entering the AI code assistant never transits foreign-controlled infrastructure.
NIS2 / NERC CIP
Critical-infrastructure isolation is met structurally by the AI code assistant, not contractually.
Zero-trust postures
No third-party endpoints to allow-list — the AI code assistant’s attack surface is your own network.
When air-gapped is the right call — and when it isn’t
Choose air-gapped when
- The network the AI code assistant must serve is already isolated — classified programs, OT networks, offline research enclaves.
- Policy prohibits any external AI API, including via proxy or private link.
- You need AI capability in disconnected field or vessel environments with intermittent or no connectivity.
Consider another mode when
- You can tolerate controlled outbound connectivity → an on-premises AI code assistant is simpler to operate and update.
- Your requirement is legal jurisdiction rather than physical isolation → the sovereign variant fits; air-gapping an AI code assistant is stricter than most regulators ask.
Same capability, different deployment mode:
AI Code Assistant: Air-Gapped vs the alternatives
| Deployment mode | Typical owner | What you gain — and give up |
|---|---|---|
| Air-Gapped (this page) | CISO / Classified Program Lead | Structural security no contract can match — in exchange for moving every model, index, and software update through an offline bundle process. |
| On-Premises | CTO / Head of Infrastructure | Maximum physical control and the strongest economics at steady volume — in exchange for owning the hardware, the capacity plan, and the upgrade cycle. |
| Self-Hosted | Platform Engineering Lead | Complete stack and model freedom with no usage meter — in exchange for your team owning operations, CVE response, and the upgrade cadence. |
| Private | CISO / Data Protection Officer | The fastest route to confidential AI — in exchange for a boundary defined by tenancy and contract rather than by a building you own. |
How to deploy an air-gapped AI code assistant
- 01
Set a latency budget and prove it on enclave hardware
Decide the acceptable p95 for inline completion before choosing a model, then benchmark candidates on the GPUs you will actually have. Adoption tracks this number more closely than it tracks completion quality.
- 02
Package models and editor extensions together
Bundle the completion and chat models, the serving stack, and the IDE extensions with a local distribution channel. Developers cannot install from a marketplace, so extension delivery has to be solved before day one.
- 03
Index the repository and tie refresh to merges
Build the retrieval index over the full codebase and internal libraries, and trigger incremental refresh from the merge pipeline so suggestions reflect the current trunk rather than last month's.
- 04
Split completion and chat onto separate capacity
Once more than a handful of developers are active, give inline completion its own GPU. Sharing one card means every long chat response degrades everyone's typing experience.
- 05
Add review agents after completion is trusted
Introduce PR-review and test-generation agents only once inline assistance is reliable. Leading with review agents attaches the tool to a critical path before developers trust it.
Where air-gapped AI code assistant projects fail
One GPU for completion and chat
A long chat generation blocks the queue and inline completion latency spikes. Developers experience this as the tool being broken, and usage drops well before anyone files a ticket.
A stale repository index
Suggestions referencing deleted APIs or old internal conventions actively slow developers down. Nightly re-indexing is not enough on an active trunk; hook it to merges.
No offline path for editor extensions
The server side is deployed and working, and developers still cannot use it because the extension only installs from an online marketplace. This is discovered late and blocks rollout entirely.
How to evaluate an air-gapped AI code assistant
Which code models run locally, and how do they benchmark on your languages?
Does context retrieval cover your monorepo or multi-repo layout?
Can it integrate with your Git platform for PR review workflows?
What telemetry, if any, leaves the developer machine?
How does per-developer cost compare to Copilot-class seats at your headcount?
Air-gapped deployments trade update convenience for structural security; budget for the offline bundle process, but the AI code assistant itself prices like any fixed in-enclave infrastructure — no meters, no per-token exposure.
An air-gapped AI code assistant, on the VDF AI platform
VDF Code delivers on-premise code assistance — local code models, repo-aware retrieval, and PR-review agents — governed like every other VDF AI workload.
Air-Gapped AI Code Assistant questions, answered
What is an air-gapped AI code assistant?
An AI code assistant provides code completion, generation, review, and refactoring to developers — and in enterprise form, does it without sending proprietary source code to an external model vendor, operating on a network with no connection to the public internet — models, updates, and telemetry all move by controlled offline transfer, so the system functions fully inside a classified or isolated enclave.
Why do enterprises choose an air-gapped AI code assistant over a cloud service?
An air-gapped AI code assistant makes no outbound calls — no license pings, no telemetry, no model API fallbacks. If a component phones home, it fails certification; the architecture must assume the internet does not exist. Air-gapped deployments trade update convenience for structural security; budget for the offline bundle process, but the AI code assistant itself prices like any fixed in-enclave infrastructure — no meters, no per-token exposure.
Which regulations drive air-gapped AI code assistant adoption?
The most common drivers are Classified handling, ITAR / export control, NIS2 / NERC CIP, Zero-trust postures. Classified handling: The AI code assistant operates inside SCIF and enclave boundaries; there is nothing to accredit outside them.
Can VDF AI run as an air-gapped AI code assistant?
Yes. VDF Code delivers on-premise code assistance — local code models, repo-aware retrieval, and PR-review agents — governed like every other VDF AI workload. VDF Code runs both the completion and review models on enclave GPUs with the repository index refreshed from your in-enclave CI, and ships the editor extensions in the same signed bundle as the server components.
Can developers use an AI code assistant on an air-gapped network?
Yes. Code-tuned open-weight models run locally on enclave GPUs and repository retrieval is built over code that is already inside the network. The parts that need deliberate planning are editor-extension distribution, which cannot rely on a marketplace, and latency, which cannot be solved by reaching for a larger hosted model.
What latency does an air-gapped code assistant need for developers to actually use it?
Inline completion needs to land in a few hundred milliseconds at p95 to feel like part of typing rather than an interruption. That constraint is what pushes air-gapped deployments toward a small quantized code model on dedicated capacity, with a larger model handling chat and review where a slower response is acceptable.
Is a self-hosted code assistant as good as a cloud one?
On raw model capability the strongest hosted models still lead, but that is only half of suggestion quality. An enclave deployment can index the entire monorepo, internal libraries, and build conventions without any exfiltration review, and better context frequently beats a stronger model working from generic training data.
How do you keep the code index current in an air-gapped deployment?
Trigger incremental re-indexing from the enclave's own CI or merge pipeline so the index tracks trunk. Scheduled nightly rebuilds drift badly on an active codebase, and stale suggestions that reference removed APIs cost developers more time than the assistant saves.
Related guides and resources
Get a migration assessment
We will map your current stack to VDF AI feature-by-feature and scope a migration path — integrations, governance, and deployment included.