VisionOps Crew

visionops_crew is a multi-agent computer vision assistant built with the Google Agent Development Kit (ADK) and Antigravity SDK. It combines model research, dataset inspection, FiftyOne/Voxel51 data curation, Hugging Face discovery, local ML engineering, Jupyter MCP notebook workflows, web search, and opt-in Antigravity SDK execution behind one orchestrated ADK app: visionops_orchestrator.

The project is designed for practical computer vision work where one request often spans several systems. For example, you can ask the orchestrator to find a dataset, import a sample into FiftyOne, inspect labels and schema quality, identify a baseline model, and produce a training plan for your local hardware. visionops_orchestrator routes each step to the right specialist and returns a single consolidated answer.

What this repository provides

• A root visionops_orchestrator ADK app that coordinates specialist agents.
• A task-planning agent for turning broad CV requests into actionable briefs.
• A CV research agent with Hugging Face MCP tools for models, datasets, Spaces, papers, documentation, and licenses.
• A data curator agent with FiftyOne/Voxel51 tools for dataset import, inspection, summaries, filtering, schema checks, and App coordination.
• An ML engineering agent with hardware profiling, local execution tools, and Jupyter MCP notebook support.
• An opt-in Antigravity SDK tool, exposed only through the orchestrator, for implementation-heavy code, image, notebook, and report generation tasks.
• A Docker Compose MongoDB service for FiftyOne metadata.
• Helper scripts for loading sample datasets, launching FiftyOne, and deleting local FiftyOne datasets.

How the system works

The main ADK app is visionops_orchestrator. It acts as a parent agent that calls specialist agents and tools, decides where each part of a request should go, compares specialist outputs, and synthesizes the final response.

flowchart LR
    user[User CV goal<br/>task, dataset, constraints] --> adk[ADK Web / App<br/>VisionOps Crew<br/>visionops_orchestrator]
    adk --> orchestrator[VisionOps Crew Orchestrator<br/>visionops_orchestrator<br/>plan, reason, route, synthesize]

    orchestrator -- Agent tool call --> planner[Vision Task Planner<br/>vision_task_planner<br/>task brief, constraints,<br/>success criteria, next step]
    orchestrator -- Agent tool call --> research[CV Research Agent<br/>cv_research<br/>models, datasets, papers,<br/>docs, licenses]
    orchestrator -- Agent tool call --> curator[Data Curator Agent<br/>data_curator<br/>local dataset import,<br/>inspection, labels, views]
    orchestrator -- Agent tool call --> ml[ML Engineer Agent<br/>ml_engineer<br/>training code, debugging,<br/>hardware, notebooks]
    orchestrator -- Search tool call --> search[Google Search Agent<br/>google_search_agent<br/>fresh web sources,<br/>official docs, APIs]
    orchestrator -- Direct callable tool --> antigravity[Antigravity Tool<br/>run_antigravity_agent<br/>SDK-powered code, image,<br/>report, artifact oversight]

    planner --> brief[Planning brief<br/>goal interpretation,<br/>task type, specialist sequence]

    research --> hf[Hugging Face MCP<br/>Streamable HTTP<br/>hub_repo_search, repo details,<br/>papers, Spaces, docs]

    curator --> direct[Direct data curator tools<br/>load_huggingface_dataset<br/>open_fiftyone_dataset]
    curator --> fiftyone_mcp[FiftyOne MCP over stdio<br/>datasets, schema, samples,<br/>labels, views, evaluation]
    direct --> fiftyone[FiftyOne datasets + App]
    fiftyone_mcp --> fiftyone
    fiftyone --> mongo[(MongoDB)]
    direct --> media[Materialized media files<br/>datasets/&lt;dataset&gt;/media]

    ml --> skillset[ADK SkillToolset<br/>hardware-profiling<br/>jupyter-notebook]
    skillset --> hardware[hardware-profiling skill<br/>CPU, RAM, frameworks,<br/>CUDA, MPS, accelerators]
    skillset --> jupyter[Jupyter MCP<br/>create, open, edit,<br/>execute, interact]
    jupyter --> notebooks[Jupyter notebooks<br/>interactive prototypes,<br/>plots, experiments]
    ml --> env[EnvironmentToolset<br/>workspace-rooted files,<br/>commands, validation]
    ml --> code_agent[ML Code Execution Agent<br/>BuiltInCodeExecutor<br/>scratch calculations and probes]
    env --> repo[Local project workspace<br/>source code, scripts, outputs]

    search --> web[Current web sources<br/>official docs, releases,<br/>vendors, APIs]

    brief --> synthesis[Unified CV answer<br/>plan, evidence, dataset insights,<br/>training steps, artifacts,<br/>caveats, next actions]
    hf --> synthesis
    fiftyone --> synthesis
    mongo --> synthesis
    media --> synthesis
    notebooks --> synthesis
    hardware --> synthesis
    repo --> synthesis
    code_agent --> synthesis
    antigravity --> synthesis
    web --> synthesis
    orchestrator --> synthesis

    classDef entry fill:#f7f4f0,stroke:#3c3c3c,stroke-width:2px,color:#1d1d1f;
    classDef orchestrator fill:#500000,stroke:#2b0000,stroke-width:3px,color:#ffffff;
    classDef agent fill:#ffffff,stroke:#500000,stroke-width:2px,color:#221f1f;
    classDef tool fill:#f7f4f0,stroke:#a7a9ac,stroke-width:1.5px,color:#221f1f;
    classDef store fill:#fff7df,stroke:#9b6a00,stroke-width:1.5px,color:#332000;
    classDef output fill:#3c3c3c,stroke:#202020,stroke-width:2px,color:#ffffff;

    class user,adk entry;
    class orchestrator orchestrator;
    class planner,research,curator,ml,search,antigravity agent;
    class brief,hf,direct,fiftyone_mcp,skillset,hardware,jupyter,env,code_agent,web tool;
    class fiftyone,mongo,media,notebooks,repo store;
    class synthesis output;

Loading

The usual flow is:

1. The user asks a computer vision question in the visionops_orchestrator ADK app.
2. The orchestrator calls the planner for broad or multi-step requests.
3. Model and dataset research go to cv_research or Google Search.
4. Local dataset work goes to data_curator.
5. Training plans, code, and hardware-sensitive recommendations go to ml_engineer.
6. Notebook prototyping goes through Jupyter MCP when the user approves notebook changes or execution.
7. Antigravity is available as an optional super-agent attached to the orchestrator. Use it when you want stronger implementation support, generated images, notebooks, reports, or artifact oversight; it is called only after an explicit user request and confirmation.
8. The orchestrator merges the specialist outputs into one practical response.

Agents

Agent name	Purpose
visionops_orchestrator	Main multi-agent orchestrator. Start here for normal use.
vision_task_planner	Plans ambiguous or multi-step computer vision workflows.
cv_research	Researches CV models, datasets, Spaces, papers, docs, and licenses through Hugging Face MCP-backed tools.
data_curator	Loads, opens, inspects, filters, and summarizes FiftyOne datasets.
ml_engineer	Builds hardware-aware training, evaluation, debugging, and notebook prototyping workflows.

run_antigravity_agent is a direct orchestrator tool, not a specialist app. Antigravity stays opt-in: the root agent calls it only when the user explicitly requests Antigravity, and file creation or editing remains denied unless the user confirms the exact write action.

Quick start

Run the local workflow in this order:

git clone git@github.com:haruiz/visionops_crew.git
cd visionops_crew
uv sync
cp .env.example .env
make run-mongo
make run-jupyter
make run-crew

Edit .env with your model settings and tokens before starting Jupyter. Then open the visionops_orchestrator app in the ADK web UI.

Configuration

Requirements:

• Python 3.11 or newer
• uv
• Docker with Docker Compose
• A Google ADK-compatible model configuration
• Optional: HF_TOKEN for private Hugging Face datasets or higher API limits

Copy .env.example to .env, then set only the values you need to override:

cp .env.example .env

VISIONOPS_ORCHESTRATOR_MODEL=gemini-3.5-flash
GOOGLE_SEARCH_MODEL=gemini-3.5-flash
VISION_TASK_PLANNER_MODEL=gemini-3.5-flash
CV_RESEARCH_MODEL=gemini-3.5-flash
DATA_CURATOR_MODEL=gemini-3.5-flash
ML_ENGINEER_MODEL=gemini-3.5-flash
ML_CODE_EXECUTION_MODEL=gemini-3.5-flash
GEMINI_API_KEY=YOUR_GEMINI_API_KEY
HF_TOKEN=YOUR_HUGGINGFACE_TOKEN
VISIONOPS_CREW_DATASETS_DIR=./datasets
FIFTYONE_APP_ADDRESS=0.0.0.0
FIFTYONE_APP_PORT=5151
JUPYTER_URL=http://localhost:4040
JUPYTER_TOKEN=YOUR_JUPYTER_TOKEN

Commands

Task	Command
Sync dependencies	uv sync
Start MongoDB	make run-mongo
Stop MongoDB	make mongo-down
Restart MongoDB	make mongo-restart
Reset MongoDB data	make mongo-reset
Follow MongoDB logs	make mongo-logs
Start JupyterLab for Jupyter MCP	make run-jupyter
Start the ADK web UI	make run-crew

Repository layout

.
|-- src/
|   `-- visionops_crew/
|       `-- agents/
|           |-- visionops_orchestrator/
|           |-- vision_task_planner/
|           |-- cv_research/
|           |-- data_curator/
|           `-- ml_engineer/
|-- scripts/
|-- assets/
|-- latex/
|-- docker-compose.yml
|-- Makefile
|-- pyproject.toml
`-- uv.lock

Troubleshooting

• MongoDB does not start: make sure Docker is running, then run make mongo-logs. If the local data directory is incompatible with the current MongoDB image, run make mongo-reset.
• FiftyOne cannot connect: start MongoDB first with make run-mongo.
• Port 5151 is busy: set FIFTYONE_APP_PORT=5152 in .env.
• Hugging Face access fails: set HF_TOKEN in .env or authenticate with the Hugging Face CLI.

Cite

If you use VisionOps Crew in research, teaching, demos, or derivative work, please cite the repository:

@software{ruiz_visionops_crew_2026,
  author = {Ruiz, Henry},
  title = {VisionOps Crew: A Multi-Agent Computer Vision Assistant},
  year = {2026},
  url = {https://github.com/haruiz/visionops_crew},
  version = {0.1.0}
}

Notes

• MongoDB is required because Voxel51 FiftyOne uses it for dataset storage, metadata, and management. Local MongoDB data is stored in docker-data/mongo.
• Antigravity SDK access belongs to visionops_orchestrator, not the ml_engineer specialist.
• The data curator and Jupyter tools inherit this project's .env configuration.