Engineering and CAD AI agents act on real geometry. Not a chatbot that describes a bracket — an agent that creates the parametric model, mates the assembly, generates the drawing, runs the check, and leaves a versioned trail of everything it did. Software Depo builds Model Context Protocol servers that expose SolidWorks and AutoCAD to AI agents through a governed, audited tool layer.
CAD actions are consequential and often hard to reverse, so the architecture is deliberate: the MCP server sits as a security boundary between the untrusted AI client and the CAD kernel. Every tool is individually scoped to least privilege, every call is recorded, and every change lands in a reproducible design history.
Both systems described below are built and working. The capabilities on this page are implemented, not planned — we are happy to walk an engineering team through either one and show the tool surface, the policy layer, and the audit trail in operation.
SolidWorks AI MCP
Exposes SolidWorks part modeling, assembly, drawing, and simulation to an AI agent through scoped MCP tools. The agent holds design intent — a prompt and a spec — and invokes only the CAD operations policy allows.
- Part modeling — sketches, extrudes, revolves, and feature edits against live
.SLDPRTgeometry. - Assembly — insert components, apply mates, manage structure in
.SLDASM. - Drawing — generate 2D views, dimensions, and annotations into
.SLDDRW. - Simulation — run FEA and design checks, returning pass/fail results and reports.
- Design history and audit — every feature, parameter, and revision versioned, traceable, and reversible.
This turns generative design into something usable: “design a bracket for a 200 N load, minimise mass” produces an editable parametric model rather than a paragraph. Natural-language edits — add a 5 mm fillet to all external edges, switch this to aluminium — execute against the live model, and simulation-in-the-loop lets the agent propose, test, and iterate with every step logged.
AutoCAD AI MCP
Drives AutoCAD from natural language instead of manual commands or one-off scripts, producing real DWG and DXF drawings. The tool surface is grouped into six individually permissioned clusters, and the server supports batch execution with dependency graphs so a multi-step drawing operation runs as one coherent unit.
- Drawing & geometry — lines, arcs, circles, polylines, rectangles, ellipses, hatches.
- Entity modify — move, copy, rotate, scale, mirror, offset, array, fillet, chamfer, erase.
- Layers & blocks — create and modify layers, freeze/thaw, lock, block insert and attribute management.
- Annotation — text, mtext, linear/aligned/angular/radius dimensions, leaders.
- P&ID symbols — ISA 5.1 symbol library, pumps, valves, tanks, instruments, process lines, flow arrows, equipment tags.
- Drawing automation — orthographic, section and detail views, intelligent dimensioning, title blocks, sheet layouts.
Dual backend and a QA gate
Operations execute either through live AutoCAD (COM / AutoLISP) for interactive sessions, or a headless DXF backend for server-side generation without a GUI. Before anything is exported, drawings pass a QA stage: geometry cleanup detects gaps, overshoots, undershoots, and misaligned angles, and a .dws standards check flags violations of company layer, dimension, and text styles. Issues route back to the agent for correction rather than reaching a reviewer.
Outputs include DWG, DXF, PDF, PNG, BOM, and JSON. Every command, batch, and revision is recorded, and operations execute as single-undo atomic units, so any AI-made change can be rolled back cleanly.
The Pattern Behind Both
The CAD specifics differ; the safety architecture does not. Both follow the discipline we apply to every production AI agent:
- Least-privilege tool scoping — each CAD operation is its own permitted or denied tool, never a general “run anything” escape hatch.
- Read-heavy by default — the agent queries geometry and properties freely; modifications and simulations go through policy.
- Everything audited — a record per call, so you can reconstruct exactly what the agent did and why.
- Reversible by design — versioned history and atomic operations mean a bad change is undone, not archaeology.
- Human review where it matters — the agent drafts and proposes; an engineer approves what ships.
Also in Development
The two MCP servers above are finished systems. The three agents below are working prototypes, actively refined — genuine builds we use and develop, but not products you can buy today. We keep the distinction explicit so you always know which is which.
AutoCAD-to-Revit Conversion Agent
Converts 2D AutoCAD floor plans into 3D Revit models: analysing DWG layers and geometry, identifying walls, doors, windows, rooms, equipment and symbols, creating the corresponding Revit elements, and selecting or creating the families to do it. It aims to preserve every meaningful object from the source drawing, detect omissions and geometry conflicts, generate validation reports, and compare the output model back against the original.
Be sceptical of anyone promising one-click 2D-to-3D conversion — drawing sets are too inconsistent for that to be honest. This is built as a closed loop: convert, validate, report what did not survive the conversion, and let a person resolve it. Successful procedures and recurring errors are stored as reusable skills, so the pipeline improves on the drawings it has already seen.
Building Metadata Agent
Identifies building objects in AutoCAD or Revit and attaches structured metadata to them — classifying HVAC, controls, electrical, security, fire-alarm and architectural objects, then applying manufacturer, model, system, room, device and point information. It validates what is missing and exports structured building data.
This is quietly the most useful of the three, because that structured output is what downstream work actually needs: estimation, commissioning, asset management, and digital-twin models all depend on knowing what each object is, not just where it sits on a drawing.
SolidWorks Design and Drawing Repair Agent
A checking-and-repair workflow for mechanical parts, assemblies, and engineering drawings: dimensions and callouts, drawing-view placement, title-block formatting, clipping and overlap detection, assembly completeness, and comparison against customer or assignment requirements. Lessons and fixes persist as reusable automation scripts rather than being re-solved each time.
Where This Pays Off
The strongest fit is where drafting and modelling work is repetitive, standards-bound, and currently a bottleneck: parametric part families and configured assemblies, P&ID generation and equipment schedules, title-block and dimensioning automation, drawing standards compliance, BOM extraction, and as-built or redline workflows. AEC, manufacturing, and process-engineering teams tend to feel this first.
Worth being clear about the limits: this is assisted, verified engineering automation — not one-click perfection. The agent accelerates the work and enforces the standards; an engineer still owns the design.
Because both servers are already built, the fastest way to judge the fit is to see them run against a workflow you recognise. Tell us which drawing or modelling task is costing you the most time and we will walk you through what an agent can safely take on — and what should stay with your engineers.