AI Coding Assistants Integration

Introduction

The CERIT-SC AI infrastructure exposes Large Language Models (LLMs) through standard API protocols, allowing you to integrate powerful AI assistance directly into your local development environment. By connecting your tools to our backend, you can leverage high-performance models (such as qwen3-coder or gpt-oss-120b) for coding tasks without needing to run them on your own hardware or rely on external commercial providers.

This chapter describes how to configure two popular tools to communicate with our API.

Integrate with Visual Studio Code

The AI chatbot can be integrated with Visual Studio Code. For this purpose, a 3rd part extension Continue is used. With this extension, you can engage AI in miscellaneous roles that can help you while coding in Visual Studio Code. The roles include simple chat, agent mode, or autocomplete. While the chat can provide a familiar conversation, the agent mode lets you analyze or edit files opened in your project. Finally, the autocomplete role suggests code as you write.

Install the Continue extension

While Visual Studio Code is running, open the Extensions tab (Ctrl+Shift+X) and search for Continue. Click on the extension and then click on the Install button. After the installation is complete, you can access the Continue extension by clicking on the Continue icon in the left sidebar.

Configure the Continue extension

First, the configuration file config.yaml must be edited. Access the Continue extension within Visual Studio Code using the icon in the left sidebar.

1. Click Open settings in the top-right corner of the Continue extension window.
2. Click Configs.
3. Click Open configuration icon at the end of the line Local Config.
4. Once the config.yaml is opened, use the following configuration with your own <api-key> (guide above):

%YAML 1.1
---
name: Local Assistant
version: 1.0.0
schema: v1
model_defaults: &model_defaults
  provider: openai
  apiKey: <api-key>
  apiBase: https://llm.ai.e-infra.cz/v1
models:
  - name: autocomplete-coder
    <<: *model_defaults
    model: qwen3-coder
    promptTemplates:
      autocomplete: '<|fim_prefix|>{{{ prefix }}}<|fim_suffix|>{{{ suffix }}}<|fim_middle|>'
    autocompleteOptions:
      transform: false
    defaultCompletionOptions:
      temperature: 0.6
      maxTokens: 512
    roles:
      - autocomplete
  - name: chat-coder    
    <<: *model_defaults
    model: qwen3-coder
    env:
      useLegacyCompletionsEndpoint: false
    roles:
      - chat
      - edit
context:
  - provider: code
  - provider: docs
  - provider: diff
  - provider: terminal
  - provider: problems
  - provider: folder
  - provider: codebase

5. Save the file. The new configuration should apply immediately.
6. Check that the FIM autocomplete is used by looking at the Continue button in the Visual Studio Code status bar (in the bottom-right corner). It should display Continue. It should not display Continue (NE). If Continue (NE) is shown, press this button and select the choice Use FIM autocomplete over Next Edit.

Usage of AI in Visual Studio Code

• The chat can be accessed by pressing the Continue icon in the left sidebar.
• In the chat, you can engage the agent mode by asking to analyze, explain, or edit the file in the currently open project. This action will require additional permissions, such as read/write permissions to the related files. You need to grant the necessary permissions to get the agent to do the job.
• Autocomplete feature will continuously suggest new code as you write. Once you see a suggestion, pressing Tab accepts the suggestion. The responsiveness of the suggestions depends on the model’s speed. You can change the model qwen3-coder in the autocomplete config section to gpt-oss-120b to get a bit faster responses, but the default model performs better in coding tasks.

Caveats

Disable all other Visual Studio Code extensions that provide AI autocomplete features. Otherwise, the Continue extension may not work properly.

Claude Code

You can deploy Claude Code and configure it to work with our models by pointing it to our API endpoint.

Installation

First, install Claude Code for your operating system by following the official instructions in the upstream repository:

• Claude Code – Getting Started

Make sure the claude CLI is available in your $PATH after installation.

Configuration

Claude Code is configured using environment variables. Export the following variables in your shell:

export ANTHROPIC_BASE_URL="https://llm.ai.e-infra.cz/"
export ANTHROPIC_AUTH_TOKEN="sk-..."
export ANTHROPIC_MODEL="qwen3-coder"
export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen3-coder"
export ANTHROPIC_DEFAULT_SONNET_MODEL="qwen3-coder"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="gpt-oss-120b"
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

Alternatively, you can define these environment variables in the settings file ~/.claude/settings.json:

{
  "permissions": {
    "defaultMode": "acceptEdits"
  },
  "env": {
    "ANTHROPIC_BASE_URL": "https://llm.ai.e-infra.cz/",
    "ANTHROPIC_AUTH_TOKEN": "sk-...",
    "ANTHROPIC_MODEL": "qwen3-coder",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "qwen3-coder",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "qwen3-coder",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "gpt-oss-120b",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

Variable description:

• ANTHROPIC_BASE_URL – Base URL of our LLM API.
• ANTHROPIC_AUTH_TOKEN – Your API token obtained from https://chat.ai.e-infra.cz.
• ANTHROPIC_MODEL – Default model to use when running Claude Code.
• ANTHROPIC_DEFAULT_OPUS_MODEL – Default model to use when running Claude Code for reasoning and complex tasks.
• ANTHROPIC_DEFAULT_SONNET_MODEL – Default model to use when running Claude Code for reasoning and less complex tasks.
• ANTHROPIC_DEFAULT_HAIKU_MODEL – Default model to use for simple tasks.
• CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC – Disables to send telemetry and various reporting (not used with non-anthropics API).

Running Claude Code

Once the environment variables are set, start Claude Code with:

claude --model qwen3-coder

You should now be able to interact with Claude Code using our backend and selected model.