NewbimAI's open-source CADViewer-Skill connects DWG drawing parsing, block counting and visual previews to agents. It does not solve the problem of “letting AI see drawings”; it enables AI to read engineering data, with humans then checking the graphics. We believe this is closer to the real needs of engineering software than relying solely on multimodal image recognition.

What this is

CADViewer-Skill is an open-source toolkit for engineering drawings, comprising a zero-installation Windows desktop app and a lightweight Skill package that can be connected to agent frameworks such as WorkBuddy and OpenCode. It can read DWG and DXF files, identify layers, standard blocks and dynamic blocks, count components, and export JSON, PNG, DXF and vector PDF files.

A Skill here is a set of specialized tools and workflows that an agent can invoke. Once connected, users issue instructions in natural language, and the agent can launch a local parsing service, obtain structured drawing data and open preview images directly in the conversation interface.

One project example is a 24-meter-by-15-meter outpatient building second-floor plan with 11 layers and 24 block categories. Of particular note is the ability to reconstruct anonymous blocks based on their parent blocks and insertion points without missing repeated insertions of the same block. This offers greater practical value for cost estimation and quantity surveying than screenshot recognition alone.

Industry view

We believe CADViewer-Skill's key advantage is that it does not separate “parsing results” from “human review.” Engineering professionals need to verify data through layer filtering, measurement and drawing comparisons, while agents handle reading and aggregation. Both use the same underlying capabilities. This approach is more restrained than recreating AutoCAD and better aligned with the data entry points required for current AI engineering deployments.

The risks are equally clear. The project currently supports only Windows and does not natively support TArch custom objects; TArch drawings must first be converted to the T3 format. Testing has also focused on architectural, MEP and construction drawings, so it cannot be assumed that mechanical models or complex industrial drawings will work with the same reliability. The open-source project's claims about counting accuracy and runtime performance still require validation against more real-world business files. More importantly, correct parsing does not mean that measurement or cost-estimation results are ready for direct use in quotations. Block definitions, version conversion and human review remain part of the chain of responsibility.

Impact on regular people

For enterprise IT: Manufacturing, construction and engineering services companies can bring DWG files into agent knowledge workflows, reducing the need for staff to manually search through drawings and perform basic counts. Before deployment, however, companies should validate their commonly used drawing formats, component rules and software versions rather than treating a prototype demonstration as a system ready for direct costing.

For individual professionals: Jobs in cost estimation, construction and project management will encounter the “agent reads drawings, human verifies drawings” workflow earlier. Its near-term value lies primarily in review and aggregation rather than replacing professional sign-off. People who can define engineering standards and identify statistical anomalies will become more valuable.

For consumers: Consumers will not immediately see mature products as a result, but decoration quotations, property renovations and home-improvement reviews may gain an additional layer of verifiable structured records. Whether this reduces pricing discrepancies will depend on the accuracy of the data, consistency of the rules and whether all parties are willing to work from the same drawings.