Preparing an AI coding agent for an air-gapped network
Prepare local models, runtimes, packages, project files, and an end-to-end test before using an AI coding agent inside an isolated network.
Using an AI coding agent in an air-gapped network requires a local model, a prepared code runtime, and the packages the work needs. Hosting the model alone does not make every development task possible. The agent may be able to write code while a build or test fails because dependencies are unavailable.
Wissly Code uses a local LLM on an internal server to write and edit code and execute it in an isolated runtime. With the required components imported in advance, requests, code execution, and app previews can run on the internal server.
Prepare four separate parts
| Component | What to prepare | What to check internally |
|---|---|---|
| Local model | Model files, serving method, usage conditions | Requests without external model calls |
| Runtime | Runtime image and language tools | Execution, builds, and tests |
| Packages | Libraries and dependencies for the task | Installation or execution without external downloads |
| Project | Code, development documents, test inputs | File reading, edits, and result review |
Wissly Code's default runtime supports Node.js and Python. Additional tools, such as Java, need to be included in the runtime image when required. Provide an internal package source or preinstall packages needed for the planned tasks.
Check model and runtime connections independently
A local model does not prevent development tools from requesting packages, remote fonts, or external APIs. Check whether the chosen workflow assumes internet access.
For example, if your model-serving setup uses Ollama, its local-only configuration is relevant. That is a configuration example for a model tool; project packages and application network requests still need separate checks.
With external access blocked, process a request, run the application, and open its preview. Classify failures as model requests, dependency downloads, or application data requests so the right component can be prepared.
Choose a first task with few dependencies
A small calculator or CSV aggregation script can make a useful initial test because inputs and expected results are clear. Starting with database connections and several internal systems can make it difficult to distinguish an agent problem from missing integration prerequisites.
For example, ask for a tool that counts sample CSV records by owner and displays the result. Supply the file format, column names, expected output, and error conditions. Check the behavior with test files before using business data.
Define runtime access and resources
Specify what the execution environment can read, write, connect to, and consume. A container does not establish every operating condition by itself. Docker's resource controls illustrate why CPU and memory settings also deserve review.
Wissly Code reads and edits files in the selected project folder. Review the runtime settings and access conditions for your deployment, and agree how long-running or resource-intensive work is stopped and restarted.
Run one complete task after import
- Start a request with the local model.
- Read project files and development notes.
- Ask for the required code.
- Check execution, build, and test results.
- Inspect changed files and the application preview.
- Confirm that another task can start in the same environment.
Review the changed code and output. The agent can prepare deployment settings and documentation, but actual production deployment follows review through your internal CI/CD system.
Information to bring to a beta discussion
Wissly Code is currently available in beta. Prepare the network conditions, development languages, project formats, required packages, and first task when discussing access.
For existing systems, see the legacy-project pilot guide. For business-tool creation, the CSV workflow guide gives a more detailed example.
The screenshots in this article show the Korean interface.