10 Aug Private Agentic Data Analytics with OpenCode, Qwen and Omniscope

In my previous article, “Proving a full European AI and data analytics stack with Omniscope”, I described a stack using European AI models, private model serving, European hosting options and Omniscope as the data, analytics and verification layer. That test used AI inside the application, mainly through Insight Explorer. In “Omniscope as a tool for AI agents”, we described another route: an external agent calling Omniscope as a toolbox for real data work. I have now tested that second route. The distinction is simple. Last time, the model was helping somebody explore a project inside Omniscope.

This time, OpenCode was the AI agent outside Omniscope. It connected through the Omniscope MCP server, discovered the available tools and authored the project itself. That changes the role of the model quite a bit.

The model is hosted in Europe; the work stays in the stack

The aim was to run a complete, private, European-hosted setup, with no data or intellectual property sent to an external AI SaaS service. Qwen3.6-27B-FP8 is the model used in this test, I needed a powerful local model. The model was hosted on a Verda GPU rig in Europe. OpenCode ran in WSL on my local Windows workstation. Omniscope ran in an isolated Hyper-V virtual machine.

So there are two separate statements here: Qwen is not a European model, but this was a European deployment of the model. The important boundary is where the model is hosted, what it can access and where the data goes.

In this setup, the data and IP stay inside the controlled environment. The architecture looks like this:

Qwen3.6-27B-FP8
Verda GPU hosting in Europe
        |
        | model inference API
        v
OpenCode in WSL
local Windows workstation
        |
        | authenticated MCP over HTTPS
        v
Omniscope 2026.2
isolated Hyper-V VM
        |
        v
Workflow, reports, insights
and answers that can be inspected

Qwen provides the model capability. OpenCode manages the agent loop, context and tool calls, whereas Omniscope does the data work and stores the resulting analytical artefacts.

The model does not need direct access to the Omniscope machine and Omniscope does not need to be exposed to the public internet. OpenCode connects to the project through an authenticated MCP endpoint, while the project remains inside the controlled environment.

This is the bit that tends to get lost in conversations about local AI. It is not enough to say that a model is open or that a GPU is running somewhere in Europe. The useful question is whether the whole working path – model, endpoint, agent, data, analytical logic and output – can operate within a boundary that the organisation controls.

Two ways to put AI around Omniscope

The earlier setup and this one are complementary. With AI inside Omniscope, a user opens Insight Explorer, asks a question and receives an answer, charts and an explanation. The model helps explore an existing project. Omniscope provides the data, workflow and verification around the answer. With an external agent, the starting point can be much less finished. The agent can be given a broader instruction and use Omniscope to find the data, create blocks, prepare calculations, build a report and inspect the result. The first is a person using AI within the analytics application. The second is an AI agent operating the analytics application. This article is about the second.

Omniscope becomes the agent’s workbench

OpenCode did not control Omniscope by clicking around a browser or interpreting screenshots. The MCP server exposed a structured set of tools for the things Omniscope already knows how to do: find data, inspect block types, add blocks, configure options, execute workflows, arrange the canvas and inspect the resulting project. The agent discovered those tools and called them in sequence. Omniscope performed the actual data operations and created normal project assets. That separation is healthy. Qwen is not pretending to be a complete analytics product outside the system. It is choosing actions from a defined toolbox. Omniscope applies those actions within its project structure and permissions. The LLM plans. Omniscope does the work.

OpenCode reports what it discovered from the analysis executed in Omniscope.

A small test of the complete path

The practical test used the Bond prices demo data. It was deliberately modest; the purpose was to exercise the full path rather than produce a grand financial analysis. From one natural-language request, OpenCode selected the data source, created the analytical blocks, configured the calculations, built the dashboards, created a multi-tab Insight Explorer and executed the workflow. No human touched the Omniscope interface during authoring. The resulting project contained grouped metrics, credit-rating analysis, correlations, coupon distributions, dashboards and an Insight Explorer connected to the analytical outputs. The findings came from executed Omniscope blocks and remained attached to the workflow that produced them. That is enough for the example: a sentence became a working Omniscope project.

The workflow and report structure created by the agent.

The agent had to make real changes

This was not a carefully staged demonstration where every tool call was known in advance. The agent used tool discovery to find the available data and inspect the schemas for the relevant Omniscope blocks. It added and wired the blocks, configured their options, executed the pipeline, arranged the canvas and inspected the final state. There were a couple of wrong turns. An early report arrangement did not produce the intended views, so the agent replaced it with instant dashboards. An initial Insight Explorer arrangement did not match the requested report structure, so it changed the design to one report with the required tabs. The corrections were expressed as requirements. Nobody had to explain which internal API call would fix the project. OpenCode selected the next tools, applied the changes and continued until the Omniscope artefact matched the request. That is the practical meaning of agentic authoring: the agent interprets an objective, chooses tools, makes changes, observes the result and corrects its own work within the permissions it has been given.

The answer stays attached to the evidence

This is where Omniscope matters most. An answer from a general chatbot can be fluent and still be difficult to audit. You can ask where a number came from, but the response may only be another generated explanation. Here, the answer has a route back to the artefact that produced it. A person can open the source data, follow the workflow, inspect the transformations and calculations, review the charts and examine the Insight Explorer output. The report is connected to the analytical logic. The insight is connected to the data preparation. The answer is connected to the project. Omniscope is therefore both the authoring environment and the verification environment. The AI agent creates the work there; people can open it afterwards, question it, edit it or continue building on it. Another AI can review the same project. That is a much better arrangement than a private chatbot alone. Privacy keeps the data inside the boundary. The Omniscope workflow makes the result visible inside that boundary.

The Insight Explorer report initialised with five questions and its first answer.

The continuation

The earlier article showed that a full European AI and data analytics stack is achievable: European-built analytics software, private model serving, European hosting options and outputs that can be inspected. This test adds the agentic layer.

The model can change, the GPU configuration can change, the agent runtime can change.
Omniscope remains the workbench where data is prepared, analytical workflows are authored, reports are built and answers are checked.

The model provides reasoning, the agent provides orchestration, Omniscope provides the controlled environment in which analytical work becomes real, persistent and inspectable.

The complete path is now:

Natural-language objective
        |
        v
AI agent plans and selects tools
        |
        v
MCP calls Omniscope
        |
        v
Omniscope authors and executes the workflow
        |
        v
Reports, insights and answers are produced
        |
        v
The project remains open to inspection

That is the step forward: from AI inside an analytics application to an AI agent that can operate the application itself, while the data, workflow and answer remain within a private European-hosted stack.

AI is planning. Omniscope is doing the verifiable work.