In the past, it was almost impossible for an amateur who had never learned any CAD software to create a 3D model out of thin air. But now you only need to tell the AI the part style you imagined in your head, and it will write the 3D model on its own, and it can also be sent directly to the 3D printer. This scene, which was only seen in science fiction movies before, was recently officially brought to the table by an open source project authorized by MIT. Recently, there is a “text-to-cad” open source project focusing on “turning coding agents into CAD engineers”, which can do things that were previously unimaginable.
The open source tool text-to-cad turns Claude Code and Codex into CAD engineers
One prompt replaces six tools
In the past, to build a robot, you had to switch back and forth between CAD software, URDF export tools, kinematic solvers, viewers, and slicing software. The entire pipeline was often stuck for days; text-to-cad took all these steps away, as long as the agent could understand a sentence of English.
text-to-cad is not a single model, but a set of “agent skill” packages. After installation, a complete set of CAD and robot tool chains will be hung in coding agents such as Claude Code, Codex, and Cursor:
- direct output STEP、STL、3MF、DXF、GLB and other industry standard formats, and can also generate robot description files. URDF / SRDF / SDF。
- built-in G-code slice and Bambu Lab print task is sent Capability, with Bambu Lab MQTT API (API of Bambu 3D printer), you can go from model to actual printing.
- supply CAD Viewer local preview,step.parts Find off-the-shelf commercial parts (screws, bearings, motors, connectors) and even experimental ones Implicit CAD, modeled on the fly in the browser using GLSL’s signed-distance field.
Complete set of toolsCompletely offline, no backend server:agent writes in models/*.py The build123d script is then used by the local Python skill to regenerate the geometry, output the artifact, and finally throw it to the viewer written in React/Vite for preview. This is critical for hardware teams that care about intellectual property rights and cannot leave design drafts in the cloud.
In the 22-second demo video, a prompt directly generates the complete design of the 7-DOF robotic arm, including kinematics and customized GUI. The entire process is driven almost exclusively by prompt words (with GPT-5.5).
Vibe coding a robot with GPT 5.5!
This is a URDF of a 7dof robot arm with functional kinematics, a custom gui, and STEP parts/assembly, 100% generated in Codex (minus the gripper).
A similar result would have taken me weeks stitching half a dozen tools together. Insane stuff. pic.twitter.com/hfikjNf1J5
— Jake Fitzgerald (@earthtojake) April 23, 2026
build123d supports the entire technology stack
The bottom layer of text-to-cad is used build123d, a set of parametric boundary representation (BREP) modeling framework written in Python, the bottom layer is connected to the OpenCascade geometry core. In other words, the agent writes solid geometry that can be read by CAD software, simulators, and CAM tool chains, not a mesh that looks like it.
This level of technology choice also explains why text-to-cad can enter the field of robotics: URDF / SRDF / SDF are the native formats of robot frameworks such as ROS, MoveIt 2, and Gazebo. Build123d directly spits out these files, and the model can be connected to kinematic planning, collision detection, and physical simulation without having to manually change the hands.
An engineer’s real reaction: It’s amazing but not quite right yet
Hacker News discussion threadBelow, the comments are roughly divided into two groups. Amazing people think this is the first decent implementation of “text-to-physical-object” after text-to-image. Someone is there LinkedIn On Threads, I have used it with Claude Opus 4.7 to iterate the design, from the atomizer power adapter to the PCB connector. “Throwing a bunch of caliper measurements and reference photos in, I can already come up with a usable skeleton.”
The skeptics grabbed the benchmark from the same repo. When some users actually tested the L-bracket, they found that the gusset (reinforcement piece) of the benchmark would overlap with the screw holes, making parts that could not be assembled at all. The joint limits of another 7-DoF arm benchmark were questioned, “LLM is easy to hallucinate on the joint limits, and it will be dangerous to use it to control the physical arm.” The more fundamental question is:Using words to describe CAD is inherently awkward., without a visual dictionary, it would be difficult for the user to clearly explain “that bevel needs to be reduced by 2mm”.
By softservo at Hacker News respondIt was explained at the time: The benchmark prompt on the homepage is just an excerpt. The actual work order prompt uses GPT 5.5 Pro to expand a sentence into complete specifications. Therefore, when users generally input vague instructions such as “help me make a fixture”, harness itself will expand it first and then feed it to the agent. “It is not intentionally misleading.” He also added that the home page prompt has been modified into a more complete version.
text-to-cad open source project website
The starting point for open source agent hardware design
Looking at the larger context, text-to-cad does not intend to become “AI CAD software” alone. Its real role is Coding agent is a low-level skill library for entering the hardware world.. In the past, AI agents mostly stayed on the screen: writing programs, writing files, and calling APIs. text-to-cad extends this chain to entities: prompt → build123d Python → STEP/URDF → G-code → Bambu Lab. Printing out the finished product is equivalent to giving a coding agent the ability to “create a physical object from scratch.”
Judging from the current user feedback, text-to-cad is more suitable in the short term. Rapid prototyping, proof of concept, there is still a distance from engineering parts that can be directly shipped; the accuracy of benchmarks, spatial coordinate consistency, and error correction during iterations all rely on users to repeatedly throw back the drawings and let the agent check itself. But for hardware innovation, robot education, and customized tool developers, just “generating 3D printable parts in one sentence” is enough to reduce the original days of work to a few hours.
As coding agents such as Claude Code and Codex continue to become stronger, coupled with the mature geometry cores of build123d and OpenCascade, the development of the open source project “AI Directly Designs and Prints Solid Parts” in the next few months is worth continuing to track.
Source: KOCPC Chinese