# Roboflow MCP for AI Agents AI Agent Connect > Roboflow MCP for AI Agents. Manage computer vision workflows like uploading images, training models, and auditing datasets directly from your AI client. It bridges the gap between your data and your agent, making it easy to handle the entire CV lifecycle without switching tabs. ## Overview - **Category:** developer-tools - **Price:** Free - **Endpoint:** https://edge.vinkius.com/vk_preview_3PcAB3lxgco6m9RlryrdB51L1KU7rDlnqGHOwOaP/ai-agent-connect - **Tags:** computer-vision, dataset-management, model-training, image-annotation, machine-learning, workflow-automation ## Description Roboflow connects to your AI agent to manage your entire computer vision pipeline through natural language. Instead of jumping between browser tabs and terminal windows to handle your data, you can now manage your projects, upload images, and monitor model performance directly within your chat interface. This Connector handles everything from dataset organization to model training and inference. You can pull images from URLs, copy existing projects to your workspace, and check your training metrics without any manual clicking. It brings your machine learning workflow into a single conversation, making it much easier to audit your data and verify how your models behave in real-time. It's one of the many specialized connectors hosted on the Vinkius catalog, designed to make your agent a power user of the Roboflow platform. Engineers can check training runs while product managers look at inference results, keeping everyone in the same loop. You can quickly see which images are missing labels, start new training runs on specific versions, or even move projects into folders to keep your workspace organized. It removes the friction of manual data management, letting you focus on refining your models rather than clicking through menus. You can also run inference on images to see how your model handles new data or check the health of your dataset to find class imbalances before you waste resources on a training run. ## Tools ### add_projects_to_folder Move projects into specific folders within your workspace. This is useful for organizing large-scale enterprise environments. ### auto_label Start an auto-labeling job using foundation models. This helps you label massive datasets much faster than doing it manually. ### cancel_training Stop an active training job immediately. Use this if you notice a model is diverging or you need to restart with different parameters. ### create_annotation_job Assign a batch of images to a human labeler or reviewer. This helps manage your manual labeling workflow efficiently. ### create_project Start a new project within your workspace. You can set up the initial metadata and categories in one go. ### delete_images Remove multiple images from a project at once. This is great for cleaning up bad data or duplicates. ### delete_project Move a project or a specific version to the trash. It keeps your workspace clean without permanent deletion. ### download_dataset Get a direct download link for a zipped dataset in formats like COCO or YOLO. This makes it easy to move data to your local machine. ### fork_universe_project Copy a public project from Roboflow Universe into your own workspace. It's a fast way to start from a pre-built model. ### get_async_task Track the status of long-running operations like large exports or forking. This lets you know when your task is actually finished. ### get_image Pull the specific details and metadata for a single image. Use this to inspect individual data points in your project. ### get_project Fetch all details, metadata, and versions for a specific project. It gives your agent the full context of your work. ### get_root Verify your API key and find your default workspace. This is the first step to ensuring your connection is active. ### get_training_results Pull the status and metrics for a specific training run. You can see mAP, precision, and recall directly in your chat. ### get_version Retrieve the metadata for a specific dataset version. This is helpful when you need to know exactly what's in a specific release. ### list_folders See all project folders in your workspace. This is a required step for managing organized enterprise environments. ### list_trash View the items currently in your workspace trash. You can see what you've deleted recently to ensure nothing was lost. ### list_workspace_projects Get a list of all projects and their info in a workspace. This gives you a high-level view of your entire portfolio. ### manage_image_tags Add, remove, or update tags on your images. This helps you organize your data for specific search queries. ### restore_trash Bring an item back from the trash to your workspace. Use this if you accidentally deleted a project or version. ### run_inference Send an image to a hosted model to see what it detects. This is perfect for quick testing of your model's behavior. ### search_project_images Filter and find specific images within a single project. This helps you find specific examples of objects or errors. ### search_workspace_images Search through all images across your entire workspace. It's the fastest way to find specific data points across multiple projects. ### start_training Begin a training run on a specific dataset version. You can trigger this via natural language to start your model development. ### upload_annotation Attach a new annotation file to an image that already exists. This is how you update labels for specific images. ### upload_image Send a new image to a project in your workspace. You can provide a URL or a Base64 string to get the data into Roboflow. ### create_folder Create a new project folder in your workspace. It helps keep your enterprise projects neatly categorized. ### get_dataset_health Check your class distribution and find missing annotations. It helps you spot imbalances in your training data quickly. ### stop_training End an active training job early. Use this to save on compute costs if the model isn't improving as expected. ## Prompt Examples **Prompt:** ``` List all the projects in my 'industrial-safety' workspace. ``` **Response:** ``` I've retrieved the projects for 'industrial-safety'. You have 3 active projects: | Project Name | Type | Status | | :--- | :--- | :--- | | Hard Hat Detection | Object Detection | Active | | Glove Compliance | Classification | Active | | Forklift Tracking | Object Detection | Active | Which one would you like to inspect? ``` **Prompt:** ``` Upload this image URL to the 'Hard Hat Detection' project. ``` **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... **Model Performance Metrics:** * **mAP:** 88.5% * **Precision:** 91.2% * **Recall:** 84.7% The training run is completed. Would you like to see the full metrics breakdown? ``` ## Capabilities ### Upload images via URLs Add new data to your projects using direct links or Base64 strings. ### Start auto-labeling jobs Use foundation models to automatically label your datasets and save time. ### Track training progress Get live updates on your model's performance metrics like recall and precision. ### Audit dataset health Check for class imbalances and missing annotations in your current versions. ### Run live inference Test your hosted models on specific images to see how they perform in real-time. ### Organize project folders Create and manage folders to keep your enterprise workspace tidy. ## Use Cases ### Auditing dataset balance An ML engineer needs to check if a dataset is balanced. They ask the agent to check the class distribution using get_dataset_health. ### Starting a training run A data scientist wants to quickly start a training run. They tell the agent to start training on version 5 of the Forklift Tracking project. ### Testing model behavior A developer needs to test a model's accuracy on a new image. They provide a URL and ask the agent to run_inference to see the results. ### Organizing workspace projects A team needs to move a project to a specific folder for organization. They ask the agent to use add_projects_to_folder to organize their workspace. ## Benefits - Stop switching tabs between your IDE and the Roboflow web UI by using upload_image and create_project from your agent. - Get instant feedback on model performance using get_training_results to see mAP and precision without opening a browser. - Clean up messy data faster with delete_images and manage_image_tags to keep your training sets accurate. - Quickly audit your data balance using get_dataset_health to spot class imbalances before you waste money on training. - Speed up your development by using fork_universe_project to pull in pre-built models into your workspace. - Automate the tedious labeling process by triggering auto_label jobs with foundation models via simple text commands. ## How It Works The bottom line is you get a direct way to manage your Roboflow account inside your favorite AI chat. 1. Subscribe to the Roboflow MCP on Vinkius and grab your Private API Key. 2. Plug that key into your AI client's configuration. 3. Ask your agent to perform tasks like "list my projects" or "start a training run" to see it work. ## Frequently Asked Questions **Can the Roboflow MCP help me manage my images?** Yes, it lets you upload images via URLs, search through your workspace, and add tags to organize your data. **How do I check my model's performance?** You can ask your agent to pull the training results for any specific version, which includes metrics like mAP, precision, and recall. **Can I use the Roboflow MCP to start training?** Yes, you can tell your agent to start a training run on a specific dataset version just by describing it. **Does the Roboflow MCP support auto-labeling?** It allows you to kick off auto-labeling jobs using foundation models to speed up your data preparation. **Can I organize my projects into folders?** Yes, if you have an enterprise account, you can use the Connector to create folders and move projects into them. **How do I find a specific image in my project?** You can ask your agent to search and filter images within a specific project or your entire workspace. **Can I download my datasets through the Connector?** Yes, the agent can retrieve a direct download link for a zipped dataset in formats like COCO or YOLO. **How can I verify if my Roboflow API key is correctly configured?** You can use the `get_root` tool. It will attempt to authenticate with your key and return the default workspace details if successful. **Can I get the training performance metrics for a specific model version?** Yes! Use the `get_training_results` tool by providing the workspace, project, and version ID. It returns mAP, precision, recall, and other training metrics. **Is it possible to export my dataset to a specific format like YOLOv5?** Absolutely. Use the `download_dataset` tool and specify the `format` parameter (e.g., 'yolov5pytorch') to receive a download link for your zipped dataset.