fovux-mcp 4.1.4

Local-first edge-AI computer vision workbench MCP server

Fovux MCP

From dataset to deployed ONNX, in one conversation.

Fovux is a professional-grade, open-source edge-AI computer vision workbench. It lets a computer vision practitioner run the full YOLO lifecycle through natural-language conversation with any MCP-compatible AI client: dataset curation, training, evaluation, error analysis, quantization, export, on-device benchmarking, and RTSP inference.

Brand: Fovux is the region of the retina responsible for sharp central vision. We help you see your models clearly.

Why Fovux?

	Fovux	Ultralytics Platform	GongRzhe/YOLO-MCP
Local-first, no account	✅	❌	✅
Full lifecycle (train→deploy)	✅	✅	❌
Error analysis	✅	Partial	❌
INT8 quantization report	✅	❌	❌
VS Code companion	✅	❌	❌
RTSP live inference	✅	❌	❌
Open source	✅	❌	✅

Status

Current distribution is source-based from this repository. Packaged releases will be published separately.

Install From Source

git clone https://github.com/oaslananka/fovux
cd fovux/fovux-mcp
pip install uv
uv sync --locked --extra dev

Quick start (5 minutes)

See docs/getting-started.md for the full tutorial.

# 1. Install from source
git clone https://github.com/oaslananka/fovux
cd fovux/fovux-mcp
pip install uv
uv sync --locked --extra dev
uv run fovux-mcp doctor

# 2. Configure your MCP client (example: Cursor / Windsurf / VS Code)
# Add to your MCP client settings:
#   "fovux": { "command": "fovux-mcp" }

# 3. Start chatting
# "Inspect my dataset at ~/data/coco128"
# "Train yolov8n on it for 50 epochs"
# "Run error analysis on the best checkpoint"
# "Export to ONNX and benchmark on CPU"

For Studio or HTTP demos, start the local transport explicitly:

uv run fovux-mcp serve --http --tcp --metrics

MCP client configuration

Cursor / Windsurf

{
  "mcpServers": {
    "fovux": {
      "command": "fovux-mcp",
      "env": {
        "FOVUX_HOME": "~/.fovux"
      }
    }
  }
}

VS Code (with MCP extension)

{
  "mcp.servers": {
    "fovux": {
      "command": "fovux-mcp"
    }
  }
}

Codex CLI

codex --mcp fovux-mcp

The tool set

Fovux MCP currently exposes 28 local tools.

Tool	Purpose
annotation_quality_check	Find common YOLO annotation quality issues.
benchmark_latency	Measure p50/p95/p99 inference latency.
dataset_convert	Convert datasets between supported formats.
dataset_find_duplicates	Detect duplicate images with perceptual hashing.
dataset_inspect	Inspect dataset structure, classes, and samples.
dataset_split	Create train/val/test splits.
dataset_validate	Validate dataset integrity and label ranges.
eval_compare	Compare evaluation outputs.
eval_error_analysis	Extract worst false-positive and false-negative samples.
eval_per_class	Report per-class validation metrics.
eval_run	Run validation for a checkpoint.
export_onnx	Export checkpoints to ONNX.
export_tflite	Export checkpoints to TFLite.
fovux_doctor	Report local environment health.
infer_batch	Run batch inference over image folders.
infer_image	Run single-image inference.
infer_rtsp	Run live RTSP inference with reconnect handling.
model_list	List local checkpoints and exports.
model_profile	Profile model size and complexity.
quantize_int8	Create INT8 quantized artifacts.
quantize_report	Compare quantized model quality.
run_compare	Compare local training runs.
run_delete	Delete non-running runs safely.
run_tag	Edit local run tags.
train_resume	Resume a stopped or failed run.
train_start	Start detached YOLO training.
train_status	Read current run status and metrics.
train_stop	Stop a running training process.

VS Code companion

Use Fovux Studio in this repo for visual run dashboards, dataset inspection, and an export wizard.

Documentation

Docs source lives in fovux-mcp/docs. Generated site/ output is a build artifact and is not committed.

uv run mkdocs build --strict

Contributing

See CONTRIBUTING.md. All contributions welcome.

License

Apache-2.0. See LICENSE and NOTICE for third-party licenses (including Ultralytics AGPL-3.0 usage terms).