Current state of AI software tools

In the German state I live in, NRW, we have the academic infrastructure provider KI:connect. This provider provides access to LLMs for all academies of the state, by aggregating requests and anonymously proxying all requests together to a single Azure OpenAI endpoint.

This is generally very good and helps to integrate various tools without paying a lot for research. Specifically, besides of limited quota of gpt-5.2 from OpenAI, there is unlimited access to openai-gpt-oss-120b and mistralai-mistral-small-4-119b-2603.

I did integrate this smoothly into various open-source software tools:

• Open-WebUI
• continue.dev as Open-Source VSCodium integration
• OpenClaw
• PicoClaw as light weight claw alternative
• Kai 9000 as private chat agent on the phone
• Jan AI as desktop chat agent

In this post I am talking about the integration to OpenCode as a coding agent

OpenCode

OpenCode is quite similar to how Cursor or Claude CLI tools are working, but it uses a BYOK (bring your own key) approach. Besides of the API-Key from KI Connect, one needes a configuration of OpenCode in ~/.config/opencode/opencode.json like this:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "kiconnect/openai-gpt-oss-120b",
  "autoupdate": false,
  "server": {
    "port": 4096,
  },
  "provider": {
    "kiconnect": {
      "name": "KIConnect NRW"
      "options": {
        "baseURL": "https://chat.kiconnect.nrw/api/v1/"
      },
      "models": {
        "openai-gpt-oss-120b": {
          "name": "GPT OSS 120B",
          "cost": {
            "input": 0.04,
            "output": 0.18,
          },
          "limit": {
            "context": 131072,
            "output": 8192,
          },
        },
        "mistralai-mistral-small-4-119b-2603": {
          "name": "Mistral Small 4",
          "cost": {
            "input": 0.15,
            "output": 0.60,
          },
          "limit": {
            "context": 262144,
            "output": 64000,
          },
        },
        "gpt-5.2": {
          "name": "GPT 5.2",
          "cost": {
            "input": 0.375,
            "output": 2.25,
            "cache_read": 0.0375,
            "cache_write": 0.01
          },
          "limit": {
            "context": 400000,
            "output": 128000,
          },
        }
      }
    }
  }
}

It is important that the context information is provided to opencode as well. Costs are given in $ per 1M token which is taken from other hosting providers which provide access to the open weights models.

This is usable for small projects and works quite well for smaller software projects.

Comparison

In contrast to Cursor Agent, Claude or Github Copilot, the tools are often not using as many tools. Therefore, they return quite fast and did not work thoroughly, evaluate different paths or do something else. This is even the case when using GPT-5.2 via API in OpenCode.

This leads me to use other tools much more. There are some opencode extensions like superpowers, but I did not find a good approach to get them to be equally good as cursor with compose-2.5 for example.

Yet, I am sure that this will improve in the next months as newer models appear. And it is of course good to be prepared. As soon as more complex tasks can be automatically solved with open-source coding agents, this will really be a game changer for privacy as well.