Three Open-Source AI Workflows: CAD, a Code-Rendered Music Video, and Chat Assistance

“AI agents can do real work” is too broad to evaluate. A better test asks what artifact an agent produces, whether a person can edit it, what the software must read or execute, and who approves the final action. Three recent open-source projects make those questions concrete: text-to-cad connects a coding agent to local CAD, pdoom-video exposes the source of a code-rendered music video, and Jev assists at the point of replying to a chat. This guide compares their workflows and gives a small, testable pilot for each.

Three different handoffs, not one generic “agent”

text-to-cad lives in an engineering toolchain. Its inputs are geometric constraints and its outputs are model files. pdoom-video lives in a creative production pipeline, turning a treatment and timing data into deterministic frames. Jev sits inside a communication workflow and leaves the final message to a person. The distinction matters because a successful demo in one pipeline says little about readiness in another.

The project overview below shows the agent-to-CAD relationship. It does not imply that a generated model has passed engineering review.

Figure 1. The project page connects an agent, local CAD tools, and exportable model files.

text-to-cad: test one measurable part before an ambitious assembly

The official README describes local model generation, a viewer, engineering drawings, and manufacturing checks. Supported agent hosts include Claude Code, Codex, and Cursor. Read the installation section for your host before running commands: the project recommends its latest release branch, requires uv for CAD, and warns against loading both its plugin and its standalone Skills in the same app. The first run may download the CAD runtime.

A good first task is an L-bracket with explicit dimensions:

Create a parametric L-bracket in millimeters. Base: 80 x 40 x 5.
Upright: 45 high, 5 thick. Add two 6.5 mm through-holes in the base;
place their centers 12 mm from the left and right edges. Add a 4 mm
inside fillet. Keep editable source. Export STEP and STL. Report the
coordinate origin, hole-center coordinates, thinnest wall, and any
fit dimensions that still need human confirmation.

This gives you objective checks. Measure the base, holes and center spacing in the viewer; open the STEP file in your usual CAD tool; inspect whether the body is valid and closed. If it will be printed, inspect material, orientation and wall thickness in a slicer. A plausible-looking preview is not a tolerance inspection. Automated DFM or DfAM checks are aids to review, not a safety certification.

The Skills list and colorful gear assembly illustrate the project's range. Use the former to identify the module you need and the latter to understand the style of output—not as proof that a moving assembly is production-ready.

Figure 2. The Skills list separates modeling, drawings and manufacturing tasks.

Figure 3. A gear example illustrates geometry, while fit and meshing still need separate validation.

Three key frames show the model at different points in its motion. Animation helps people see the parts and spot obvious intersections, but it does not validate load or fatigue.

Figure 4. Initial state: identify the moving parts.

Figure 5. Mid-motion state: look for interference.

Figure 6. Later state: motion remains a visual check, not a structural test.

One operational setting deserves attention. The README says text-to-cad sends anonymous usage and crash telemetry by default, excluding user files, paths and prompts; uvx cadgen telemetry off disables it. On a managed device, also review runtime downloads and local execution permissions before installing a plugin.

pdoom-video: study the method behind one video, not a one-click generator

pdoom-video is the source of a particular music video. Its repository splits the visual treatment, renderer and scene API, word-level lyric timing (data/lyrics.json), beat and section data (data/audio.json), timeline and scene modules. The author says Claude Code assisted with the concept, alignment, analysis, renderer and scenes. That makes the repository a useful case study in creative engineering. It does not mean any song can be dropped in and converted automatically.

Figure 7. The directory structure reveals the separation between audio analysis, creative documentation and rendering code.

For a practical reading route, start with TREATMENT.md: what does each scene contribute to the song? Then compare data/lyrics.json and data/audio.json to see how words and beats become coordinates in time. Finally inspect app/src/timeline.ts to see when scenes enter and leave. That sequence yields a transferable production pattern: treatment → timing data → scene code → export.

The README's local preview path is cd app, bun install, then bunx vite. The offline renderer additionally uses Chrome and ffmpeg. A full high-quality 4K export is far more demanding than a browser preview. For a first pilot, change one scene's copy, color or camera parameter, check the preview, then render a short test. The code is MIT-licensed, but the repository explicitly excludes the song and lyrics from that code license. Reusing the workflow for your own publication requires separately cleared audio and other assets.

A storyboard template for a code-rendered clip

These intervals are a reusable exercise, not a transcription of the video's exact edit.

Jev: draft beside the conversation, but leave sending to the user

The current Jev README is broader than an Android-only description. Its Chinese Android edition lists QQ, Feishu, X and WhatsApp; Windows and Apple Silicon macOS have separate projects. An iOS keyboard exists as source to build yourself, rather than an App Store release. Check Products and Releases before installation because platform status can change.

Figure 8. The project page now points to several platform-specific variants.

Figure 9. Read the platform matrix as a current implementation snapshot, not a permanent compatibility guarantee.

On Android, Jev uses accessibility and overlay permissions to read the current chat, infer intent and risk, generate candidate replies, and fill the one you select. You still press send. Desktop versions use window capture and on-device OCR or text APIs; iOS uses copied messages. These paths have different access and data implications.

Figure 10. The overlay offers drafts while the user remains responsible for sending.

Figure 11. Confirm the configured provider and key before trying a real chat.

Figure 12. Separate reading, model inference, drafting and human approval during review.

Test it with a synthetic conversation that contains no personal information. Let one participant make an ambiguous request; ask whether the assistant seeks clarification, changes tone without changing facts, and avoids promising an unapproved deadline. Then inspect exactly what context is sent to the configured provider. The project says Android chat text goes to the model provider selected by the user, not through a Jev server. That is not the same as “the conversation stays on-device.” Customer, contractual, medical, financial and child-related chats need an explicit organizational decision before use. The repository also warns that integrating with third-party chat apps may conflict with their terms.

Code0: turn each pilot into a repeatable task specification

Teams comparing model calls through Code0 can keep separate evaluation sets for these three task types instead of assuming one model setup solves every integration. For CAD, record dimensions, exported files and manual measurements. For video, keep the treatment, short render and rendering time. For chat, use anonymized test conversations and record whether a draft makes an unauthorized promise. Compare cost per accepted artifact, including retries and human editing. text-to-cad inherits its host agent's model setup; Jev's custom provider configuration must be checked separately. This guide does not claim that any of these repositories natively integrates with Code0.

Choose by artifact, permissions and rework

Hardware prototyping teams should pilot text-to-cad with a small measured part and inspect the output in their existing CAD software. Motion designers can study pdoom-video's treatment and timeline, then render a short scene. People seeking better wording in chat may benefit from Jev's “judge, draft, human send” sequence, but must treat the conversation content as sensitive input.

Keep one trial log for all three: input, artifact, minutes of human rework, and actions the tool must never perform automatically. Run the same task several times and record failures, not just a best-case screenshot. A project belongs in a regular workflow only when its output is inspectable, its permissions are acceptable and the handoff back to a person is clear.

Sources

• text-to-cad official repository and installation guide
• pdoom-video official repository, renderer documentation and asset-license notes
• Jev official repository, platform status and privacy notes

FAQ

Can I send a STEP file from text-to-cad straight to a factory?

Do not skip engineering review. Inspect dimensions, assembly fit and tolerances in a CAD system and run checks appropriate to the manufacturing process. An openable file is only a starting point.

Can pdoom-video generate a new music video from another song automatically?

No. It is the implementation of one work. A new song needs new timing analysis, treatment, scenes and rights-cleared audio.

Does Jev send chat messages on its own?

The project says it fills a candidate reply and the user presses send. Still verify the behavior, permissions and chat-app terms for the version you install.

Can all three projects use the same model API endpoint?

Do not assume so. text-to-cad inherits its host agent's model setup, pdoom-video is a specific creative codebase, and Jev uses a user-configured provider. Check each project's configuration and run a limited integration test.