English · Español

Genix Apolo CAD

An agent-native 3D parametric CAD for industrial machinery.
Built to be driven by an AI agent (Claude Opus or others) over MCP — or by hand in your browser.

What it is

Apolo is a headless parametric CAD for designing real machines. Its edge isn't the kernel (it uses OpenCascade, like FreeCAD) but its agent-native architecture:

Every operation is a command against an API. The whole document is an editable command log. And the same JSON Schemas that generate the UI also generate the agent's tools. One single source of truth.

The upshot: an AI agent can design complete machines end to end —not just autocomplete— and verify them: detect interferences, simulate gravity, look at a render (vision) and emit fabricable shop drawings. It's STEP-interoperable, drivable by a human or an AI, and usable as a headless backend that other tools/agents call.

MVP vertical: conveyors / material handling.

What makes it different

• 🤖 Genuinely agent-native. Not a chatbot bolted onto a CAD: the agent is a first-class client of the same API as the UI. It designs, measures, validates and fixes on its own.
• 🧾 Document = command log (event-sourced). Geometry is never stored → KB-sized files, free undo/redo, parametric editing of any past command.
• 🧬 Schema-driven. Adding a command to the registry makes it appear in the toolbar, the dialogs, the properties panel and the agent's tools — without touching anything else.
• 🔢 Variables and expressions. Any numeric field accepts "=expression" (=width-2*profile). Changing a variable regenerates the whole model.
• 🎯 Declarative edge/face selectors (by direction, face, length, proximity) — goodbye to the fragile topological naming problem.
• 🧱 Machine templates = super-commands (e.g. create_belt_conveyor, create_take_up): they inherit parametric editing, undo, BOM and agent exposure for free.

How to use it

▶ Via an AI agent (MCP) — the primary way

You mostly operate Apolo by talking to an agent. It exposes 79 MCP tools, so any MCP-compatible client (Claude Code, Claude Desktop, etc.) running a capable model —Claude Opus or others— can design entire machines. The repo ships a .mcp.json:

{
  "mcpServers": {
    "apolo-cad": {
      "command": ".venv/Scripts/python.exe",
      "args": ["-m", "apolo.mcp_server"],
      "env": { "APOLO_URL": "http://127.0.0.1:8000" }
    }
  }
}

With the server up, you ask your agent in plain language:

"Design a 4 m × 600 mm belt conveyor for 1–15 kg parcels, with a hollow-shaft gearmotor and gravity-type take-up tensioning. Check there are no interferences and show me a render."

And the agent:

1. Models with run_batch (atomic batches: one regenerate, one undo step), referencing parts from the same batch with $k and dimensions with =expression.
2. Perceives with render_view (returns an image → vision), get_topology, measure.
3. Validates with check_interference, engineering_check, gravity_test (simulates what falls).
4. Documents with drawing / drawing_set / assembly_manual → shop drawings, cut lists, BOMs and step-by-step assembly manuals.

The write core is minimal (run_command / run_batch / edit_command + undo/redo + set_variable) and covers the entire command registry — there is no tool per command. The rest of the 79 tools are for reading, perception, drawings and validation. Everything the agent does lives in the log: editable, undoable and reproducible. Changes show up live in the browser.

🖱 By hand (web UI)

A three.js viewport with PBR materials, shadows and a ViewCube; a schema-driven ribbon (Create / Sketch / Modify / Assemble / Library / Robotics); a parametric properties panel; "pro CAD" shortcuts (move/rotate with snap, isolate, fit, measure, section). The agent and the UI are two equal clients of the same API: what one does, the other sees.

Gallery

Belt conveyor — maintenance lift, a declared motion study	Folding door — wood + translucent glass

Everything above is produced by the engine itself: shaded renders and animations come from render_view / motion.gif (VTK) — the very images the agent looks at to review its own work. The animation is a named motion study: the agent declares the joint keyframes, then scans the whole travel for collisions.

Architecture

   AI agent (MCP) ───┐        ┌── React + three.js (web UI)        equal clients
                     ▼        ▼                                    of the same API
                core/apolo/api      FastAPI · REST + WebSocket
                     │
                     ▼
   doc        document = command log (event-sourced · undo/redo · KB-sized .apolo)
   commands   command registry + JSON Schemas  (single source of truth)
   kernel     build123d / OpenCascade  (B-rep geometry, render, measure, picking)
   library    catalog (231 refs) · BOM · machine super-commands
   assembly   joints · mates · constraints · connectivity / gravity
   drawing    pro 2D drawings  (HLR → SVG/DXF/PDF · sections · dimensions · drawing sets)
   physics    gravity / stability  (MuJoCo)

Clean, non-negotiable boundaries: kernel (pure geometry) ⟂ commands/registry (operations + schemas) ⟂ doc (log/state) ⟂ api (transport) ⟂ agent/mcp (AI clients) ⟂ ui. Designed to scale (many commands, modules and clients).

Capabilities

• Modeling — primitives, fillet/chamfer/shell/drill, patterns, mirror, revolve, extrude, sweep/loft (incl. closed loops and helix), sheet metal with flat-pattern DXF/SVG export, constrained 2D sketching (in-house scipy solver), STEP import.
• Assembly & kinematics — persistent face mates (re-solved on edit), joints (fixed/revolute/continuous/prismatic), rail and N-DOF constraints, motion study (animate the joints and scan collisions along the path).
• Library & BOM — a 231-reference catalog populated from real standard dimensions (ISO/ASTM/DIN/EN: bearings, profiles, fasteners, joinery, hardware…) + super-commands (create_belt_conveyor, create_weldment, create_frame, create_sheet_metal, create_take_up, create_drive_roller, robot arm). BOM with cut list and CSV export.
• Engineering validation — engineering_check (vertical rules: belt speed, motor torque, support…), check_interference (OCCT booleans), and gravity-based assembly validation (declare joints/grounds and simulate what falls with convex hulls in MuJoCo).
• PRO manufacturing drawings — HLR projections → SVG/DXF/PDF, dimensions with arrows and tolerances, A-A/B-B sections with per-material hatching, detail views, title block + revisions, full drawing sets, automatic hole dimensioning, exploded views, light GD&T, step-by-step assembly manuals, and an Inventor-style color shaded iso. All from a declarative spec (drawing(spec)) the agent composes.
• FEA (linear static) — per part (tet mesh + safety factor from σ_vm) and bonded multi-material assembly: the whole welded frame under design load, with FS reported per piece and deflection contrasted against the analytical check.
• Engineering deliverables — this is where Apolo beats what a CAD ships by default: a calculation report (A4 PDF with design basis, formula, substitution and safety factor per check, each citing its standard — CEMA/ISO/DIN/EN/AISC), a costed BOM and quotation, tolerance stack-up (worst-case + RSS over ISO 2768 / ISO 286), an installation sheet (anchor load per support) and a delivery check — a green/amber/red gate that refuses to call a design done while it still floats, collides or lacks declared fastening.
• AI — MCP server (79 tools), vision rendering, agent session memory, chat auto mode (execute → verify → fix).

Requirements

• Python 3.11–3.13 (with OCP/build123d binary wheels)
• Node.js 18+
• (Optional) An Anthropic API key (ANTHROPIC_API_KEY) for the AI assistant embedded in the UI

Installation

# Python core
python -m venv .venv
.venv\Scripts\python -m pip install -e core
.venv\Scripts\python -m pip install pytest httpx   # for the tests

# UI
cd ui
npm install
npm run build    # builds ui/dist, served by the server itself

Running it

$env:ANTHROPIC_API_KEY = "sk-ant-..."   # optional (UI's AI assistant)
.venv\Scripts\python -m uvicorn apolo.api.main:app --port 8000

Open http://localhost:8000. To connect an agent over MCP, keep the server running and point your MCP client at the repo's .mcp.json. Optional env vars: APOLO_MODEL (default claude-opus-4-8), APOLO_DB (SQLite path).

Tests

.venv\Scripts\python -m pytest tests -q   # 1355 tests