オフライン環境での Vibe Coding: VS Code Chat / Agent をローカル SLM/LLM で使うために整理したことThis article walks through the practical steps and caveats for using local…

Running language models locally has become one of the more practical ways to bring AI-assisted coding into environments where sending source code to a cloud service is not an option. A recent Qiita post examines exactly this scenario: wiring a local small or large language model (SLM/LLM) into Visual Studio Code's Chat and Agent features so that "vibe coding" can continue without an internet connection. For teams in regulated industries or on air-gapped networks the appeal is obvious, but the author's central message is that the reality is more involved than downloading a single model file.

The premise the article pushes back on is a familiar one. When developers discuss embedding generative AI into their toolchains, someone usually suggests that a local LLM can replicate most of the cloud experience offline. The author agrees the direction is sound but cautions that moving from that idea to a working setup on real machines surfaces a series of configuration hurdles that are easy to underestimate.

At a minimum, an offline stack needs more than the model weights. A local inference runtime, commonly Ollama, llama.cpp, or a similar server, has to expose the model over an API that VS Code can reach. VS Code's Chat and Agent functionality, typically delivered through GitHub Copilot Chat, has gained the ability to point at alternative model providers, including locally hosted ones, but connecting them requires deliberate configuration rather than an out-of-the-box switch.

Several caveats tend to bite in genuinely disconnected environments. Some extensions still expect to authenticate or check for updates over the network, and telemetry or sign-in flows can block startup if they cannot reach their endpoints. Anyone planning an air-gapped deployment has to account for these dependencies, cache what can be cached in advance, and verify which features degrade gracefully when the network is absent.

Model choice is another axis. Small language models are attractive because they run on modest CPUs or consumer GPUs, but Agent mode places heavier demands on a model than plain chat does. Agentic workflows depend on reliable tool calling, structured output, and enough context length to hold file contents and instructions, and smaller quantized models may handle these tasks inconsistently. The article's framing of "SLM/LLM" reflects this trade-off: the right pick depends on available memory, acceptable latency, and how much autonomy the agent is expected to exercise.

The phrase "vibe coding," popularized by Andrej Karpathy, describes a loose style of programming in which a developer states intent and leans on the model to produce and revise code. Doing this offline is now more feasible thanks to the maturing ecosystem around local inference. Ollama, LM Studio, and a growing catalogue of openly licensed models such as those in the Llama, Qwen, Mistral, and Gemma families have lowered the barrier to running capable models on a single workstation. VS Code's own openness to third-party and local providers, alongside comparable moves by editors like Cursor and by the Continue extension, points to broader industry interest in giving developers control over where inference happens.

It is worth keeping expectations calibrated. A locally hosted SLM will not necessarily match a frontier cloud model on complex reasoning or long multi-step edits, and the offline setup trades that ceiling for privacy, predictability, and independence from external services. For many internal or sensitive projects that trade appears reasonable, but it should be tested against the actual tasks a team performs rather than assumed.

The practical value of the write-up lies in cataloguing these prerequisites and constraints rather than promising parity with cloud assistants. It appears aimed at engineers evaluating whether an offline setup is viable for their team, and it is likely most useful as a checklist of what to configure and what to test before committing. The honest takeaway is that offline AI coding is achievable today, but it rewards careful setup: cho