Coding assistants for air-gapped and zero-egress environments
Compare AI coding assistants for air-gapped environments. Learn how local weights, self-hosted endpoints, and zero data retention affect capabilities.
In this article7
The three architectures, and why they are not substitutes
When no prompt and no line of source code may leave the network, three architectures remain possible, and they are not substitutes. The difference between them is where the prompt goes, and it can be checked in each tool's own documentation before any product is named. The first architecture runs local weights on the developer's own machine: the prompt never leaves it. The second points the assistant at a self-hosted endpoint inside the perimeter, so the prompt crosses at most the internal network. The third keeps using a hosted service but relies on contractual limits: the prompt has already left the network, and the contract governs what happens to it afterwards.
Assistants, runtimes and models that run inside the perimeter
The assistant layer, the runtime layer and the model layer are separate decisions. The table below lists each assistant's documented architecture, offline setup, accepted backend and licence or deployment model. A dash means the page lists no value.
| Assistant | Architecture supported | Documented offline or air-gapped setup | Model backend accepted | Licence or deployment model |
|---|---|---|---|---|
| aider | Local or self-hosted endpoint | Connects to a local Ollama server via OLLAMA_API_BASE | Ollama models, including ollama_chat endpoints | Open source, terminal-based |
| Cline | Local or self-hosted endpoint | Runs against a local runtime (Ollama, LM Studio or Atomic Chat) on localhost | Ollama, LM Studio, Atomic Chat; OpenAI-compatible endpoints | Open source, IDE extension |
| Continue | Local or self-hosted endpoint | Install from the .vsix, turn off anonymous telemetry, set a local default model in config.yaml | Ollama and self-hosted models via HuggingFace TGI, vLLM, SkyPilot, Anyscale Private Endpoints or Lambda | Open source, IDE extension |
| Tabby | Self-hosted endpoint | Self-hosted server; the project documents itself as an open-source, self-hosted AI coding assistant | Coding LLMs including CodeLlama, StarCoder and CodeGen | Open source |
| Tabnine | Self-hosted in the perimeter | Private installation on-premises or in a VPC, deployable in a completely air-gapped environment; Enterprise option | Tabnine's own cluster in a private installation | Commercial, Enterprise tier |
The runtime layer is configured separately from the assistant. Ollama documents a local-only mode (OLLAMA_NO_CLOUD=1 or disable_ollama_cloud) that turns off its cloud models and web search[7]. Alternatively, vLLM serves an OpenAI-compatible endpoint on the team's own hardware[8], allowing teams to integrate it by pointing a coding assistant at your own model endpoint.
The model licence is a gating fact before anything reaches production. Qwen3-Coder-30B-A3B-Instruct[9] and Devstral Small 1.1[10] are released under Apache 2.0, per their model cards. Other models carry non-production licences, such as Codestral-22B-v0.1[11], which matters before they go into production use. Devstral Small 1.1's card states it is light enough to run on a single RTX 4090 or a Mac with 32 GB RAM, whereas choosing larger local models requires calculating what a 70B model costs in VRAM. The distinction between downloadable weights and a genuinely open licence is covered in the comparison of open-weight and open-source LLMs and their licences.
What zero data egress means in a contract and on the network
Claude Code's data-usage page is a worked example of the third architecture: zero data retention is available to qualified accounts on Claude for Enterprise, is not included in the standard Enterprise plan, and is enabled on a per-organisation basis by the account team after confirming eligibility[12]. The same page states that the WebFetch domain safety check sends the requested hostname to api.anthropic.com before fetching a URL, that this check runs regardless of which model provider is configured, and that if the network blocks api.anthropic.com, WebFetch requests fail until the domain is allowlisted or the check is disabled in settings. A network-level block is verifiable by the reader with ordinary tooling; a retention promise is verifiable only through the contract and its audit rights. The two are different controls, and while policies may dictate what EU residency can and cannot promise, the words air-gapped, offline, self-hosted and zero data retention do not mean the same thing.
Which assistant features stop working
A smaller local model gives up capability against a hosted frontier model: context handling, tool-calling reliability and multi-file agentic editing degrade as parameter count and available VRAM shrink, and Cline's own local-inference guidance recommends enabling its compact prompt mode and keeping tasks focused because smaller context means faster responses. Which features stop working is a separate question from model quality, because some features depend on a vendor service rather than on model size.
- Hosted chat and cloud agents: unavailable when the assistant's vendor service is unreachable; Claude Code's cloud sessions, for example, run in vendor-managed virtual machines.
- Web search and web fetching: Ollama's local-only mode disables web search.
- Telemetry-based updates and model downloads: Continue's offline setup requires turning off anonymous telemetry; Ollama pulls models from the internet and needs a proxy or manual transfer to reach an air-gapped network.
- Extension installs: Continue's offline guide directs users to install from a downloaded .vsix rather than a marketplace.
Choosing under a no-egress constraint
Deciding on an architecture requires clarifying the exact nature of the constraint. If the environment is fully disconnected, local weights running directly on the developer's machine are the only viable path. If the requirement restricts traffic to the corporate network, a self-hosted endpoint in the perimeter expands the range of viable models without violating the boundary. If the policy permits traffic to leave the network but forbids persistent storage by third parties, a hosted service with contractual retention limits provides the most capable models, provided the organisation accepts the audit and compliance trade-offs.
Frequently asked questions
Can an AI coding assistant work offline?
Yes. Assistants like Continue, Tabnine, and Tabby support offline or air-gapped deployments. They require local weights running on the developer's machine or a self-hosted endpoint within the network perimeter to function without internet access.
What does zero data egress mean?
Zero data egress is a network-level constraint ensuring no prompts or source code leave the corporate network. This differs from zero data retention, which allows data to leave the network but relies on a contract to ensure the provider does not store it.
Can I run a coding model on my own hardware?
Yes. Models like Devstral Small 1.1, which has 24 billion parameters, can run on a single RTX 4090 or a Mac with 32 GB RAM. Runtimes like Ollama and vLLM can serve these models on local hardware to keep data inside the network.
Which assistants support a self-hosted backend?
Assistants like aider, Cline, Continue, Tabby, and Tabnine can be configured to point to a self-hosted backend. Some, like aider, Cline and Continue, accept OpenAI-compatible endpoints served by runtimes like vLLM.
Is zero data retention the same as air-gapped?
No. Zero data retention means the data leaves your network but the provider promises not to save it, as seen with qualified accounts on Claude for Enterprise. An air-gapped setup physically or logically prevents data from leaving the network entirely.
Sources
Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score
Related reading
All Sovereignty →LLM APIs with zero data retention: A provider comparison
Compare 7 LLM inference providers that offer zero data retention. Learn what each commitment covers and how to get API logging restrictions in writing.

Who holds the keys to your inference data
Learn what BYOK inference means for European data sovereignty, how to manage encryption keys, and why zero retention offers a simpler compliance path.

Average data center size MW: enterprise vs hyperscale
Compare the average data center size in MW across enterprise, colocation, and hyperscale campuses, and learn why capacity outranks floor space.
