Skip to main content

IDE Integration

Use mittwald's GDPR-compliant AI models‍ directly in your development environment for code generation, completion, refactoring, and debugging. All models are hosted in Germany without storing user data or using it for training.

Available Models​

The service offers several AI models optimized for different development tasks. For a complete list of available models and their specifications, see the AI models documentation‍.

Recommended models:

  • DeepSeek-V4.1-Flash-1017: the largest model we offer and our recommendation for coding agents — 1,048,576-token context, image input, tool calling. It's in the Testphase‍
  • Qwen3.8-27B-NVFP4: the Stable choice for production use, 256,000-token context, up to 5 images per request
  • Qwen3.6-35B-A3B-FP8: fastest of the coding models, 256,000-token context, good for completion and quick edits
  • Kolibri-1: German-language chat and writing, 1,048,576-token context, text only. It's in the Testphase‍
  • gpt-oss-120b: text only, LTS lifecycle, 131,072-token context
  • Ministral-3-14B-Instruct-2512: small and fast, 262,144-token context, for lightweight assistants

Supported IDEs and Tools​

IDE/ToolIntegrations
JetBrains IDEs (IntelliJ, PHPStorm, ...)JetBrains AI‍, Continue‍
Visual Studio Code
CursorCline‍
WindsurfCline‍
ZedZed‍

Running an agent in the terminal instead of an editor? See CLI Agents‍ for Hermes Agent, Pi and opencode.

Quick Start​

  1. Get your AI Hosting API key: See Gaining access‍
  2. Choose your IDE: Select from the integrations above
  3. Configure endpoint: https://llm.aihosting.mittwald.de/v1
  4. Select model: DeepSeek-V4.1-Flash-1017, or Qwen3.8-27B-NVFP4 if you need a Stable model

Best Practices​

Model Selection​

  • DeepSeek-V4.1-Flash-1017: start here for agentic coding, refactoring, and tool use. Set the context window to 1,048,576 tokens and max_tokens to at least 32,768, because the model reasons before it answers. For hard reasoning and tricky debugging, set reasoning_effort: "max"
  • Qwen3.8-27B-NVFP4: production setups that need a model with the status Stable
  • Qwen3.6-35B-A3B-FP8: inline completion and quick edits, where latency matters more than depth
  • Kolibri-1: chat and writing in German. It calls tools only with tool_choice: "auto", so don't use it with agents that force a tool call
  • gpt-oss-120b: automations that need the LTS lifecycle and process text only
  • Ministral-3-14B-Instruct-2512: lightweight assistants and simple chat

Long-Context Tips​

  • Avoid sending entire repositories; retrieve relevant files/function bodies only
  • Summarize prior steps and include IDs/anchors to rehydrate context on demand
  • Stream responses; raise max_tokens only when necessary
  • Prefer deterministic sampling for tools and edits: temperature <= 0.2. Reasoning models such as DeepSeek-V4.1-Flash-1017 and Kolibri-1 are the exception: keep the sampling parameters from their model pages
  • Limit repo scope; request specific files/functions to keep context efficient
  • Prefer JSON/tool outputs when orchestrating multi-step changes

MCP Server Integration​

Many IDEs support Model Context Protocol (MCP)‍ servers for enhanced capabilities:

  • Context7: Up-to-date library documentation
  • Brave Search: Web search with privacy
  • Directus: Headless CMS management
  • OpenSearch: Search, index inspection, mappings, and cluster diagnostics
  • TYPO3: TYPO3 CMS content and backend automation through MCP
  • mittwald: Infrastructure management through natural language

For a broader ecosystem list, see the MCP Server Registry‍.

See individual IDE guides for MCP configuration details.

Next Steps​

Choose your IDE or tool to get started:

Disclaimer​

Third‑party tools, MCP servers and external links are provided for convenience without endorsement or warranty. Use them at your own risk, review licenses and privacy policies, and scope access conservatively (least privilege). Avoid sending sensitive data to external services unless required and permitted.

On this page