Skip to main content

CLI Agents

Command-line coding agents run in your terminal rather than inside an editor. They read and write files, run commands, and call tools on their own, driven by a model you point them at. Because mittwald AI Hosting speaks the OpenAI API, any agent that accepts a custom OpenAI-compatible endpoint works with it — with your code and prompts staying on infrastructure in Germany.

For agents that live inside an editor instead, see IDE Integration‍.

What an agent needs from the endpoint​

A coding agent needs three things beyond plain chat, and mittwald AI Hosting provides all three:

  • An OpenAI-compatible /v1/chat/completions endpoint. This is what the baseURL or base_url setting in each agent points at.
  • Streaming responses, so output appears as it is generated instead of arriving in one block at the end.
  • Tool calling. This is the one that matters most and the one that most often breaks against non-OpenAI endpoints. An agent without tool calling cannot read a file or run a command — it can only talk. Verified working on Qwen3.8-27B-NVFP4, Qwen3.6-35B-A3B-FP8 and gpt-oss-120b.

Examples of supported agents​

These three are documented here because they were verified end to end against mittwald AI Hosting. Any agent that lets you point it at a custom OpenAI-compatible endpoint works the same way.

AgentConfig fileMCP supportSkills
Hermes Agent‍~/.hermes/config.yamlBuilt in (hermes mcp add)Built in, plus registries
Pi‍~/.pi/agent/models.jsonBuilt in (pi mcp add), since 0.99.0Directory discovery or --skill
opencode‍~/.config/opencode/opencode.jsonBuilt in (mcp config block)~/.config/opencode/skills/

Choosing a model​

All models are listed in the AI models documentation‍. For agentic work:

Use forModelContext
Main driverDeepSeek-V4.1-Flash-10171,048,576
Main driver, StableQwen3.8-27B-NVFP4256,000
Subagents, cheap turnsQwen3.6-35B-A3B-FP8256,000
Hard reasoning, debugging (reasoning_effort: "max")DeepSeek-V4.1-Flash-10171,048,576
Routing, classificationQwen3.5-0.8B262,144
Documents and imagesGLM-OCRsee model page

Image understanding (vision) isn't limited to one entry here — DeepSeek-V4.1-Flash-1017, Qwen3.8-27B-NVFP4 and Qwen3.6-35B-A3B-FP8 all accept image input as Base64-encoded data URLs, though Qwen3.8-27B-NVFP4 caps requests at 5 images.

Start with DeepSeek-V4.1-Flash-1017 as the driver. It's the largest model we offer and the one we recommend for coding agents and long-horizon agentic work, and its context window holds a whole code base. Keep thinking on, which is its default, and leave room in the agent's output limit: the model page recommends 131,072 to 262,144 tokens for coding agents at reasoning_effort: "max", see Coding agents‍. The model is in the Testphase‍: we can shut it down at short notice, and its behavior and feature set can still change.

For a production-stable driver, use Qwen3.8-27B-NVFP4 (Stable) and leave thinking on, which is its default. With thinking off, agents on this model can get stuck repeating the same tool call — see Agent loops in non-thinking mode‍. If an agent supports per-subagent model overrides, point the subagents at Qwen3.6-35B-A3B-FP8 — the work they do is mostly search and summarising, and the faster model keeps long sessions responsive.

Qwen3.5-122B-A10B-FP8 will be disabled on 28.12.2026. If your agents still use it, move them to its successor Qwen3.8-27B-NVFP4, or to DeepSeek-V4.1-Flash-1017 if you want the largest model we offer and a model in the Testphase works for you.

Kolibri-1 isn't a good fit for CLI coding agents: it reads text only, scores lower than Qwen3.6-35B-A3B-FP8 on tool-calling benchmarks, and rejects tool_choice="required" and a named tool_choice. Use it for German-language assistants and RAG instead, see Kolibri-1‍.

Quick start​

  1. Get an AI Hosting API key from the AI-Hosting section of your mStudio project — see Gaining access‍.
  2. Pick an agent from the table above and follow its page.
  3. Configure the endpoint as https://llm.aihosting.mittwald.de/v1.
  4. Verify tool calling works before starting real work. Each guide has a verification step for this.

Keeping context under control​

Large context windows are not a reason to send more. Agents burn context fastest on file contents and command output.

  • Point the agent at specific files and functions rather than whole repositories.
  • Use the sampling parameters recommended on the model's page. Don't lower temperature to 0: greedy decoding makes the models repeat themselves.
  • Set a step limit where the agent has one, so a run that loops ends early instead of using up your token quota.
  • Raise max_tokens only when a reply is actually being truncated.
  • Where the agent supports compaction or summarisation, leave it enabled for long sessions.

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