# Replicate MCP for AI Agents AI Agent Connect > Replicate lets you run machine learning models directly from your AI client. You can trigger image generation, text processing, or audio models without setting up your own infrastructure. It handles the heavy lifting of GPU management while giving you a direct line to run predictions, manage your custom models, and browse the public library. ## Overview - **Category:** developer-tools - **Price:** Free - **Endpoint:** https://edge.vinkius.com/vk_preview_ULeVbcRjoLCkIr9x93q2HML8VJd4AXmk4CWHHPso/ai-agent-connect - **Tags:** machine-learning, ai-models, gpu-computing, inference-api, stable-diffusion ## Description Imagine you're building an app that needs to generate unique images or process heavy audio files. Usually, you'd have to spin up a server, manage CUDA drivers, and worry about scaling. This Connector changes that by letting your AI agent do the work for you. You can ask your agent to find a specific model on Replicate, run a prediction with your custom parameters, and then just grab the result. It's about moving from how do I host this to what do I want to build. You can manage your own fine-tuned models, track training jobs, and see real-time logs without leaving your chat interface. Since Vinkius hosts this Connector, you get a unified way to plug into these high-end machine learning capabilities across any of your favorite AI tools. It's the bridge between your high-level ideas and the actual compute power needed to make them real. You don't need to worry about the underlying hardware or the complexity of the inference API. You just need the output. ## Tools ### cancel_prediction Stop a prediction that is currently running. Use this if you notice a mistake in the input. ### create_deployment_prediction Run a prediction using a dedicated deployment. This ensures faster, more reliable results for production. ### create_model Create a new model entry on the Replicate platform. Use this to organize your custom assets. ### create_prediction Trigger a model to run a specific task. It sends your inputs to the model and starts the process. ### create_training Start a new training session to fine-tune a model. This is how you customize models on your own data. ### delete_model_version Remove a specific version of a model. Use this to clean up old or broken versions. ### get_account See your account details and organization info. It's the quickest way to check your current status. ### get_collection Pull details for a specific model collection. This helps you see how models are grouped. ### get_model_version Get the details and OpenAPI schema for a specific version. This lets your agent understand the exact inputs needed. ### get_prediction Check the status and final output of a prediction. Use this to see if your generation finished. ### get_training Check the current status of a training job. This tells you if your fine-tuning is still running. ### get_webhook_secret Retrieve the secret key for your webhooks. You'll need this to verify signatures from Replicate. ### list_collections See all curated collections of models. This is a good starting point for finding new tools. ### list_hardware See all available hardware SKUs. Use this to understand the compute options available to you. ### list_predictions See a history of your recent predictions. This is your go-to for looking back at previous runs. ### search_models Search for public models on Replicate. Use this to find the right tool for any ML task. ### update_model Change the metadata for an existing model. This helps you keep your custom model info current. ### create_deployment Set up a private deployment with specific autoscaling rules. This is great for production environments. ### get_model Fetch specific details about a model. This is useful for checking configuration and metadata. ### list_model_versions See every version of a specific model. This helps you pick the right one for your task. ## Prompt Examples **Prompt:** ``` Find a model that can turn text into a 3D mesh. ``` **Response:** ``` I found a few options for 3D generation. The most relevant one is 'shap-e/shap-e'. Would you like me to pull the input schema so we can try a prompt? ``` **Prompt:** ``` Start a training job for a model using my custom dataset. ``` **Response:** ``` I've initiated the training session. You can track the progress here. I'll let you know as soon as the status changes from 'starting' to 'running'. ``` **Prompt:** ``` What's the status of my last 5 predictions? ``` **Response:** ``` Here is the summary of your recent activity: | ID | Status | Result | |---|---|---| | pred_123 | Success | [Link] | | pred_124 | Running | In progress | | pred_125 | Success | [Link] | Would you like me to check the output for the finished ones? ``` ## Capabilities ### Run model predictions Execute any model version with your own custom inputs and handle the results. ### Find public models Search for existing models and their schemas to see what's available for your project. ### Create custom models Build and host your own machine learning models on the platform. ### Monitor training jobs Track the progress of your fine-tuning sessions in real-time. ### Manage private deployments Set up dedicated deployments with specific autoscaling for production use. ### List hardware options See available GPU SKUs and hardware descriptions for your compute needs. ## Use Cases ### Prompting for Art A designer wants a specific style of image. They ask the agent to find a Stable Diffusion model and run a prediction with a detailed prompt. ### Fine-tuning a Niche Model A developer needs a model trained on specific company data. They use create_training to start the process and get_training to watch it finish. ### Scaling a Production App An engineer needs to serve model results to thousands of users. They use create_deployment to set up private autoscaling. ### Batch Processing A data scientist needs to run 100 different inputs through a model. They use list_predictions to keep track of every job's status. ## Benefits - Skip the infrastructure headache by running models on Replicate's GPUs instead of your own hardware. - Find the right model for your specific task using search_models instead of browsing through endless documentation. - Get production-ready results faster by using create_deployment to handle autoscaling for your custom models. - Monitor your fine-tuning progress in real-time with get_training to see exactly how your custom data is being processed. - Save time on development by using get_model_version to automatically fetch the correct input schema for your agent. - Manage your entire ML portfolio from one place using list_predictions to track every generation and training job. ## How It Works The bottom line is you get instant access to production-grade machine learning without managing any infrastructure. 1. Paste your Replicate API token into your AI client's MCP settings. 2. Describe the machine learning task you want to complete to your agent. 3. Get the output file, image URL, or status update directly in your chat. ## Frequently Asked Questions **How does the Replicate MCP help with machine learning?** It lets your AI agent execute models directly. Instead of writing code to call an API, you just tell the agent what to generate or process, and it handles the communication with Replicate's infrastructure. **Can I use this Connector to run Stable Diffusion?** Yes, you can use search_models to find the best version and then run predictions to generate images directly from your chat. **How do I manage my custom models with the Replicate MCP?** You can use it to create, update, and delete models. Your agent can also pull specific versions and their schemas to ensure it's sending the right data. **Can my agent monitor my model training?** Absolutely. Use the get_training tool to check progress. Your agent can even give you updates as the training moves through different stages. **Is it possible to scale my model for production?** Yes, you can use create_deployment to set up a private deployment with autoscaling, which is perfect for handling high traffic. **How do I know what inputs a specific Replicate model needs?** The Connector can fetch the OpenAPI schema for any model version. This allows your agent to see exactly what parameters are required before it tries to run a prediction. **How can I check if my prediction has finished and see the output?** Use the `get_prediction` tool with your Prediction ID. It will return the current status (starting, processing, succeeded, or failed) along with the output URLs or data once completed. **Can I search for specific types of models like 'image-to-text'?** Yes! Use the `search_models` tool with your query. It will return a list of public models matching your terms, including their owners and descriptions. **Is it possible to stop a model that is taking too long to run?** Absolutely. Use the `cancel_prediction` tool with the target Prediction ID to immediately stop the execution and prevent further usage costs.