AI 代理與工程軟體
Cogram Studio Opens Headless FreeCAD via MCP, Letting Agents Create CAD and BIM Files Directly
Cogram has introduced a cloud-based modeling environment shared by humans and AI agents, using FreeCAD 1.1, the OpenCASCADE geometry kernel, and an MCP toolset to work with 3D models and technical drawings. Agents can already complete constrained, iterative modeling tasks, but engineers must still inspect complex models for geometry, dimensions, and solid constraints.
Cogram Studio transforms a traditional desktop CAD application into a remote workspace that AI agents can operate. Its backend runs FreeCAD 1.1 in headless mode, with geometric operations handled by the OpenCASCADE kernel. Workspaces can import and export STEP, IFC, STL, DXF, and FreeCAD’s native FCStd files, rather than merely generating meshes or images for presentation.
It exposes an MCP server with tools for executing FreeCAD Python scripts, reading model state, creating or modifying views, spreadsheets, and drawing sheets, and importing terrain data. Developers can have agents such as Codex and Claude Code read the official `skill.md` and then start a one-hour session, or use the built-in Pi-based agent. The browser interface handles 3D viewing, measurement, and technical drawing management, enabling humans to verify the results directly after each round of agent modifications.
The technical breakthrough is not that models have suddenly learned new geometry formats, but that agents now have access to a stateful, observable CAD execution environment that they can call repeatedly. This is closer to how a coding agent works than asking a language model to produce a complete FreeCAD script in one attempt: first create a part, then inspect the result, measure its dimensions, and correct errors. MCP also allows existing agent frameworks to enter the modeling workflow without integrating a dedicated SDK.
However, Cogram explicitly states that the system is currently best suited to highly repetitive or narrowly scoped tasks. Asking an agent to complete a complex model in a single attempt may produce something that looks plausible but fails dimensional and structural inspection. Engineering teams should also verify session isolation, script permissions, file provenance, and audit logs for model changes. The next developments to watch are whether the platform adds geometric constraint validation, simulation feedback, version diffs, and reproducible testing mechanisms, allowing agent-generated CAD to progress from interactive prototypes into formal design workflows.