Requirements

Before installing Dexter, confirm your environment meets every requirement below. Missing any one of these will block a pipeline run at a specific step.

System requirements

Requirement	Minimum	Used for
OS	macOS 12+, Linux (Ubuntu 20.04+)	Blender, OpenCode
Python	3.10+	Tool scripts, JSON validation
Blender	3.6+ (4.x recommended)	GLB measurement, assembly, rendering, USD export
RAM	8 GB	Blender Cycles rendering (4 views × 1920×1080)
Disk	~500 MB per run	Images, GLBs, blends, renders, USD
Network	Required	OpenAI, fal.ai, OpenCode API calls

Software dependencies

OpenCode (agent runtime)

Dexter has no standalone Python pipeline driver. The OpenCode orchestrator agent is the only entrypoint. You need:

OpenCode CLI installed and authenticated
A connected model via Codex OAuth (/connect in the OpenCode TUI)

curl -fsSL https://opencode.ai/install | bash
opencode --version   # verify

Python packages

From requirements.txt:

Package	Version	Purpose
`fal_client`	latest	fal.ai Hunyuan 3D image-to-mesh API
`openai`	latest	OpenAI `images.edit()` for component PNGs
`requests`	latest	HTTP downloads (GLB files from fal.ai)
`jsonschema`	latest	JSON Schema validation gate
`pyyaml`	latest	`configs/base.yaml` parsing

pip install -r requirements.txt
python3 -c "import fal_client, openai, jsonschema, yaml; print('OK')"

Blender

Blender must be callable as blender on your PATH, or set paths.blender_binary in configs/base.yaml:

blender --version
# macOS alternative:
# /Applications/Blender.app/Contents/MacOS/Blender --version

Blender is used for four pipeline steps:

blender_measure_glbs.py: raw bounding boxes per GLB
blender_assemble.py: build scene from assembly.json
blender_render_views.py: diagnostic PNG renders for the critic
blender_export_usd.py: final USD export for Isaac Sim

All Blender scripts run in background mode:

blender --background --python tool_scripts/blender_assemble.py -- --layout <path> --output <path>

Arguments after -- go to the Python script, not Blender.

API keys and accounts

Variable	Service	Required when	Get it
`OPENAI_API_KEY`	OpenAI Images API	Image generation step	platform.openai.com (opens in a new tab)
`FAL_KEY`	fal.ai Hunyuan 3D	3D mesh generation step	fal.ai (opens in a new tab)
OpenCode auth	OpenCode / Codex	All agent steps	`opencode` → `/connect`

export OPENAI_API_KEY=sk-...
export FAL_KEY=...

API keys are never stored in config files or committed to the repo. Scripts read them from the environment at runtime.

Cost estimate per run

Costs vary by part count and placement loop iterations:

Service	Approximate cost	Notes
OpenCode (analyze, critic)	Depends on plan	Subagents use your logged-in model
OpenAI Images (`gpt-image-2`)	~$0.02-0.10 per PNG	One PNG per part
fal.ai Hunyuan 3D	~$0.10-0.50 per GLB	One GLB per part
Blender	Free	Local compute

A 4-part dishwasher with 5-6 placement iterations typically costs $2-5 in API fees.

Input image requirements

Place source photos in input_images/. Supported formats: PNG, JPG, JPEG, WebP.

Guideline	Why
Single object, clear view	Analyze agent needs to see distinct moving parts
Neutral/plain background	White backgrounds work best for component image isolation
Show articulation state	If doors are open or drawers extended, placement matches that pose
≥ 512×512 resolution	1024×1024 or higher recommended

Bundled examples: input_images/dishwasher.png, refrigerator.png, washingmachine.png, oven.png.

Pre-flight checklist

Run these before your first pipeline:

# Python deps
python3 -c "import fal_client, openai, jsonschema, yaml; print('Python OK')"
 
# Blender
blender --background --version
 
# API keys
test -n "$OPENAI_API_KEY" && echo "OPENAI_API_KEY set" || echo "MISSING: OPENAI_API_KEY"
test -n "$FAL_KEY" && echo "FAL_KEY set" || echo "MISSING: FAL_KEY"
 
# OpenCode
opencode --version
 
# JSON validator
python3 tool_scripts/common.py --help

Next: Installation

Introduction Installation & Run