JsCAD – free online 2D CAD and DXF editor

Turning an image or scan into a drawing

For users and AI agents (through the MCP server). A picture is only a picture: to edit it as CAD it has to be redrawn as lines, text, hatches and blocks.

Small sheets (forms, tables, details)

1. Attach the image (IMAGEATTACH, or drop the file on the drawing) as a reference.
2. Measure positions from the image pixels rather than by eye, so table lines and cells line up.
3. Draw on separate layers (frames, grid, fills, text, logo) next to or over the image; agents use cad_batch_add_entities.
4. Compare with the original (cad_export_svg for agents) and correct the differences.

Large plans (architectural drawings, many sheets of detail)

Reading a big plan in one look does not work: fine text and dimensions are lost. Split the work: algorithms find the geometry, the AI decides what it means, and the result is checked against the image.

1. Underlay and scale first. Attach the image on a locked layer. Use SCALE with Reference (or ALIGN) so a known dimension or the scale bar measures true. One wrong scale ruins everything after it, so confirm it.
2. Work in tiles. Cut the image into overlapping tiles (about 1000–1500 px), handle each one and place the results in drawing coordinates.
3. Geometry by code, not by eye. Image tools (for example OpenCV) clean the scan (straighten, black and white, remove specks), find straight lines and arcs (door swings), merge broken pieces, make near-horizontal and near-vertical lines straight, close gaps at corners, and pair parallel thick lines into walls (MLINE). The AI reviews and fixes the result.
4. Text by OCR. OCR (for example Tesseract or PaddleOCR, both read Arabic) reads room names and dimensions; the AI decides which text is a dimension, a room name or a note.
5. Blocks and layers. Doors, windows, sanitary fittings and furniture become one block each, inserted many times. Sort everything into layers such as walls, doors, windows, text and dimensions.
6. Build in batches. Agents send a few hundred objects per cad_batch_add_entities call and use cad_execute_script for procedural parts, so thousands of coordinates do not pass through the conversation. Save progress tile by tile (cad_export_dxf).
7. Check automatically. Export SVG, overlay it on the image, score each tile for how well the lines match, and redo only the tiles that score low.
8. Ask when unsure. Hand-drawn plans, blurry scans, hatching and stamps defeat automatic tracing. Keep the image underneath as a reference and ask instead of guessing.

What JsCAD does and does not do

• JsCAD provides the drawing side: image underlay (embedded in the drawing), layers, blocks, multilines, tables, text, scaling by reference, DXF and SVG export, and the MCP tools above.
• JsCAD does not trace images or read text from them automatically: steps 3 and 4 need outside image and OCR tools.
• Expect automation to do most of a clean printed plan, with review for the rest; poor scans need much more manual work.