fovux-mcp 1.4.0

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

Packaged releases are produced by GitHub Actions in oaslananka/fovux-kit. Install fovux-mcp from PyPI when you need the signed release artifact, or use the source workflow below for development.

Install From Source

git clone https://github.com/oaslananka/fovux-kit
cd fovux-kit/fovux-mcp
uv sync --frozen --extra dev

The Apache-2.0 core keeps YOLO engine dependencies optional. Install the yolo extra only when the Ultralytics backend and its separate AGPL/commercial terms are appropriate for your use case:

uv sync --frozen --extra dev --extra yolo

Quick start (5 minutes)

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

# 1. Install from source
git clone https://github.com/oaslananka/fovux-kit
cd fovux-kit/fovux-mcp
uv sync --frozen --extra dev --extra yolo
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"
        }
    }
}

The tool set

Fovux MCP 1.3.0 currently exposes 47 local tools.

Tool	Purpose
active_learning_queue_list	List review queue entries from the SQLite database.
active_learning_queue_rank	Rank unlabeled images by uncertainty using a YOLO checkpoint and populate the review queue.
active_learning_queue_submit	Submit label corrections for a queue entry, copy the image to the dataset, and write the YOLO label file.
active_learning_select	Rank unlabeled images by model uncertainty for annotation prioritization.
annotation_quality_check	Inspect YOLO labels for common annotation mistakes before a bad dataset wastes training time.
benchmark_latency	Measure local inference latency and throughput for a model artifact.
dataset_augment	Create a local augmented YOLO dataset copy using deterministic transforms.
dataset_convert	Convert between supported YOLO and COCO dataset layouts.
dataset_find_duplicates	Perceptual hash duplicate detection for image datasets.
dataset_inspect	Comprehensive dataset statistics for YOLO or COCO exports.
dataset_split	Create reproducible train, val, and test splits.
dataset_validate	Deep integrity checks for YOLO datasets.
demo_init	Initialize a demo workspace for first-run onboarding.
deployment_advise	Analyze deployment readiness, preflight checks, parity, and benchmarks.
distill_model	Start a student-model training run with teacher-model distillation metadata.
eval_compare	Evaluate multiple checkpoints on the same dataset and rank the results.
eval_error_analysis	Inspect confusion patterns and worst examples beyond headline metrics.
eval_per_class	Return a sorted per-class view over evaluation output.
eval_run	Run a validation pass on a checkpoint.
export_onnx	Export a checkpoint to ONNX and optionally verify parity.
export_reproducibility_bundle	Export a reproducibility bundle zip file for a training run.
export_tflite	Export a checkpoint to TFLite, optionally with INT8 enabled.
fovux_doctor	Inspect the local Fovux environment before training, exporting, or opening Studio live views.
generate_support_bundle	Generate a redacted support bundle zip file containing system diagnostic information.
get_policy_status	Retrieve the current security policy status and allowed tools for the active environment.
infer_batch	Run inference over an image directory and persist the detections as a reusable manifest.
infer_ensemble	Run inference with multiple checkpoints and fuse the detections.
infer_image	Run structured inference on a single image.
infer_rtsp	Run live inference over an RTSP stream with reconnection logic.
list_audit_events	Retrieve audit event logs from the local database.
model_compare_visual	Generate visual comparison artifacts between two model checkpoints.
model_list	List tracked checkpoints and exported model artifacts.
model_profile	Profile a checkpoint so you can choose between accuracy, size, and compute cost before training or export.
quantize_int8	Produce an INT8 ONNX export using a calibration dataset.
quantize_report	Compare original and quantized checkpoints on the same evaluation set.
run_archive	Archive a completed training run to a compressed file.
run_compare	Generate a markdown and PNG summary across multiple training runs.
run_delete	Deletes a non-running training run from the SQLite registry and, by default, removes its run.
run_tag	Replaces the tag list for a training run. Tags are stored in the local SQLite registry and used by.
set_policy_mode	Set the local security policy mode to adjust permissions and confirmation prompts.
sync_to_mlflow	Sync a training run to a local or remote MLflow tracking server.
train_adjust	Adjust hyperparameters of a running training run.
train_preflight	Perform preflight checks and return a diagnostic training compatibility summary.
train_resume	Resume a stopped or failed run from its latest checkpoint.
train_start	Launch a non-blocking YOLO training subprocess.
train_status	Read the latest state and metrics for a tracked training run.
train_stop	Stop a running training subprocess and mark the run as stopped.

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

Fovux core is Apache-2.0. The Ultralytics YOLO backend is optional and carries its own AGPL/commercial licensing boundary; install the yolo extra only when that backend is appropriate for your use case. See LICENSE, NOTICE, and docs/adr/0003-ultralytics-adapter-boundary.md.