# Roboflow MCP > Roboflow manages your entire computer vision pipeline, letting you handle everything from dataset uploads to model training runs through natural language conversation. You can create new projects, download datasets in COCO or YOLO formats, and monitor metrics like mAP without leaving your agent. ## Overview - **Category:** developer-tools - **Price:** Free - **Tags:** computer-vision, dataset-management, model-training, image-annotation, machine-learning, workflow-automation ## Description Managing a computer vision workflow used to mean jumping between six different dashboards: one for data annotation, another for version control, and a third just to check performance metrics. This MCP changes that. Your agent connects directly to Roboflow, letting you manage the full CV lifecycle using simple prompts. You can upload raw images via URL or Base64, track dataset health, and even fork public projects straight into your private workspace. Whether you're an ML engineer monitoring a training run or a data scientist needing specific versions of data for custom scripts, this tool makes it happen in plain conversation. This capability is available through the Vinkius catalog, giving you one connection point to hundreds of other services. You start by setting up projects and then use your agent to run inference on images right away. ## Tools ### add_projects_to_folder Adds existing projects into a designated folder within an enterprise workspace. ### auto_label Starts an automated labeling job using foundation models to speed up annotation. ### cancel_training Stops a model training run that is currently active, saving computational resources. ### create_annotation_job Assigns a batch of images to specific labelers and reviewers for human annotation work. ### create_folder Creates new project folders within an enterprise workspace structure. ### create_project Sets up a brand new, empty machine learning project in your account. ### delete_images Removes multiple images from a specific project folder. ### delete_project Deletes an entire project or dataset version, moving it to the trash bin. ### download_dataset Generates a download link for all images in a zipped file format (COCO, YOLO, etc.). ### fork_universe_project Copies an existing public project from Roboflow Universe to your private account. ### get_async_task Tracks the status of long-running background operations, like large exports or forks. ### get_dataset_health Checks the structural integrity of a dataset, looking at class balance and missing annotations. ### get_image Retrieves specific metadata details for one image file in your project. ### get_project Fetches comprehensive details and version history for a given project. ### get_root Verifies the connection credentials and retrieves the default workspace name. ### get_training_results Pulls metrics and status updates for a specific, completed training run version. ### get_version Gets metadata details about a particular version of your dataset. ### list_folders Lists all the project folders available within an enterprise workspace. ### list_trash Displays a list of items that have been deleted and moved to trash. ### list_workspace_projects Lists all projects associated with the current workspace account. ### manage_image_tags Adds, removes, or sets descriptive tags on one or more images. ### restore_trash Brings a deleted item back from the trash bin into active use. ### run_inference Runs an immediate test on an image using pre-hosted models to see its predicted output. ### search_project_images Finds and filters images based on criteria within a single project. ### search_workspace_images Searches for images across the entire workspace using defined filters. ### start_training Initiates a model training job on a specified dataset version. ### stop_training Stops an active model training process early if it's not performing well enough. ### upload_annotation Attaches a file containing annotations to an existing image asset. ### upload_image Uploads new images directly into a specified project folder. ## Prompt Examples **Prompt:** ``` List all projects in my Roboflow workspace 'industrial-safety'. ``` **Response:** ``` I've retrieved the projects for 'industrial-safety'. You have 3 active projects: 'Hard Hat Detection' (object-detection), 'Glove Compliance' (classification), and 'Forklift Tracking'. Which one would you like to inspect? ``` **Prompt:** ``` Upload this image URL to the 'Hard Hat Detection' project in workspace 'industrial-safety'. ``` **Response:** ``` Uploading image... Success! The image has been added to the 'Hard Hat Detection' project. It is currently in the unassigned batch. Would you like to assign it to a specific split? ``` **Prompt:** ``` Show me the training metrics for version 5 of the 'Forklift Tracking' project. ``` **Response:** ``` Fetching results for version 5... The model achieved a mAP of 88.5%, with a precision of 91.2% and recall of 84.7%. The training run is completed. Would you like to see the full metrics breakdown? ``` ## Capabilities ### Build and Organize Projects Create new project folders or fork public Roboflow Universe projects into your private workspace. ### Manage Image Assets and Data Versions Upload images using URLs, manage dataset versions, check class distribution health, and download the data in required formats like COCO or YOLO. ### Train and Monitor Models Start model training runs on specific dataset versions and retrieve detailed performance metrics, including mAP, precision, and recall. ### Search and Verify Data Filter images within a project or an entire workspace to audit data quality or run real-time inference against hosted models. ## Use Cases ### Validating Model Performance Post-Deployment A product manager needs to know if the model can detect 'hard hats' on images taken in a new warehouse environment. They ask their agent, and it uses `run_inference` against a few test images, instantly providing visual confirmation of false positives or negatives. ### Expanding Dataset Scope for New Classes A data scientist realizes the 'Glove Compliance' model is weak on dirty gloves. They use `upload_image` to add a batch of new photos and then run `get_dataset_health`. The agent reports low class distribution metrics, telling them exactly where the dataset needs more focus. ### Quickly Setting Up an Enterprise Project An ML engineer joins a new team. Instead of starting from scratch, they use `fork_universe_project` to copy a proven 'industrial safety' template into the workspace. They then run `create_folder` and organize their assets immediately. ### Retraining After Data Changes A team manually corrects hundreds of annotations, which are uploaded via `upload_annotation`. Now they need to retrain. The agent uses `get_version` to lock the new data state and then runs `start_training`, giving them a predictable path from cleanup to deployment. ## Benefits - Speed up data prep by running `auto_label` jobs; your agent handles assigning foundational models to label images, saving manual annotation time. You'll get labeled assets ready for training almost instantly. - When you need to audit your data, use the built-in search tools like `search_workspace_images`. Instead of manually clicking through thousands of files, your agent pulls up exactly what you’re looking for based on tags or metadata. - Stop guessing if your model works. Run an instant test using `run_inference` to check how a hosted model behaves against new images right in the conversation window. It's immediate feedback without setup. - Data versioning is simplified: Use `get_version` and then `start_training`. You tell your agent which dataset version you want, and it handles pointing the training job to the correct data snapshot. - Organization becomes simple. Need a new project? Just run `create_project`. If you find a good public example, use `fork_universe_project` to start from that baseline instead of building from scratch. ## How It Works The bottom line is you manage complex visual data workflows without ever touching a dashboard or API call yourself. 1. Subscribe to this MCP and provide your Roboflow Private API Key. 2. Your agent uses the key to authenticate and retrieve your default workspace details. 3. You prompt your agent with a command, like 'Train a model on version 3,' and the tool executes the job and returns the status. ## Frequently Asked Questions **How does Roboflow MCP handle large datasets?** It handles them by providing tools like `get_dataset_health` to audit class distribution across massive uploads and allowing you to download structured data via `download_dataset` for external use. **Can I train a model using Roboflow MCP?** Yes, your agent initiates training runs by calling `start_training`. It then lets you monitor the process and retrieve metrics using `get_training_results` until it's complete. **What if I need to improve my model with new images?** You can use `upload_image` or `search_workspace_images` to find and add the necessary assets. After adding them, you run `get_dataset_health` again to ensure the balance is correct before retraining. **Does Roboflow MCP help with project organization?** Absolutely. You can use `create_project` or `add_projects_to_folder` to structure your work, and even clone public examples using the `fork_universe_project` tool. **How do I test my model without running a full training job?** You can run an immediate prediction check by calling `run_inference`. This tests your existing models on new images and gives you real-time results instantly.