AI connections (MCP)
Connect FrameTrain to Claude and ChatGPT
With the Model Context Protocol (MCP) an AI assistant can operate the FrameTrain app on your computer: prepare datasets, load models, start training and evaluate results. The tools always run in the app – frame-train.com only signs you in and relays calls.
Connector URL
https://frame-train.com/api/mcpHow it works
- You connect Claude or ChatGPT with the connector URL and sign in with your FrameTrain account.
- The assistant sees the tools of your FrameTrain app and calls them when you ask it to.
- frame-train.com relays each call to your open app; the app runs it locally and sends back the result.
- Long tasks such as training or downloads return a job ID right away; the assistant checks progress with status tools.
Requirements
- FrameTrain account with an active subscription or a running trial.
- FrameTrain app installed and signed in with the same account.
- In the app under Settings > AI Connections: "Allow remote access via frame-train.com" turned on.
- The app must stay open while the assistant works with FrameTrain.
Set up in Claude
- In Claude (claude.ai, desktop or mobile) open Settings > Connectors and choose "Add custom connector".
- Name: FrameTrain. URL: the connector URL above. Leave OAuth client ID and secret empty.
- Click Connect, sign in to frame-train.com and allow access.
- Enable FrameTrain in a chat from the tools menu and describe what you want to train.
Set up in ChatGPT
- In ChatGPT turn on developer mode under Settings > Apps & Connectors > Advanced (until FrameTrain is listed in the app directory).
- Choose "Create", name: FrameTrain, paste the connector URL, authentication: OAuth.
- Sign in to frame-train.com and allow access.
- Pick FrameTrain in a new chat from the plus menu.
Local: Claude Desktop, Claude Code, Cursor
On the same computer you can also connect without going through frame-train.com: in the app under Settings > AI Connections you find "Connect to Claude Desktop", the command for Claude Code and the configuration for Cursor. This local server only listens on 127.0.0.1.
Tools
Your app provides the exact list; it grows with new versions. Overview by area:
- System
- Check app version, GPU, VRAM, RAM and Python environment.
- Models
- List local models, search Hugging Face, download, import, rename and delete models.
- Datasets
- Browse datasets, preview, load from Hugging Face, import locally, split and delete.
- Dataset Builder
- Create projects, add text, images, audio and video, annotate, check quality and export as a dataset.
- Training
- Check presets and requirements, estimate resources, start, monitor and stop training.
- Synapse
- Build custom model architectures as a graph, validate them, generate code and train.
- DevTrain, Lab, Tests
- Edit training scripts, try out models, start test runs and fetch results.
- Analysis and hosting
- Compare runs, read metrics and logs, serve models locally and query them.
Approvals and safety
- Read-only tools run without asking. Actions such as deleting, overwriting files or starting training must be confirmed by you in the app each time.
- Tools only access the FrameTrain data folders or files you confirm in the app via a dialog.
- We store access tokens only as hashes. You can disconnect any connection instantly in the dashboard under AI connections or in the app.
Privacy
Training, models and datasets stay on your computer. Only tool calls and their results pass through frame-train.com. See the privacy policy (section AI connections) for details.
Troubleshooting
- The assistant says FrameTrain is not open.
- Start the app and turn on remote access under Settings > AI Connections. The dashboard shows whether the app is online.
- Only the system_status tool appears.
- Your app has never reported with remote access. Turn on remote access, wait a moment and reload the tool list in the assistant.
- A call reports a timeout.
- Usually an approval dialog is waiting in the app. Confirm it and let the assistant check the status.
- Sign-in fails.
- AI connections require an active subscription or trial. Disconnected connections must be reconnected in the assistant.
Questions? Contact us through support in the dashboard.