Neptune.ai MCP. Audit model versions and run metrics via conversation.

When you hook up your AI client to this Neptune.ai MCP Server, you get direct command access to your entire machine learning lifecycle—from project setup details right down through model performance metrics. You don't have to navigate dashboards; you just talk to your agent and it does the heavy lifting.

Project Discovery & Setup
You can start by figuring out what workspaces are available. Use list_projects to get an exhaustive list of all Neptune ML projects you've got access to. Once you know which project you wanna dig into, use get_project to pull the full metadata and configuration details for that specific workspace.

This tells you exactly how the environment is set up before you even look at a model.

Searching Historical Runs & Models
If you're auditing old work or trying to replicate a setup, your agent can search deep into history. You use search_runs to pinpoint specific ML experimentation runs within any project. This tool lets you find tracked data and performance boundaries across dozens of training attempts without guessing which run was the winner.

Once those runs are located, you'll need to know what models were packaged with them; call list_models to pull a complete inventory of every trained model that lives inside that project.

Deep Metric Extraction & Analysis
To really understand why one model is better than another, you gotta look at the raw numbers. For any run ID found via searching history, use get_attributes to retrieve precise performance metrics and parameters. This gives you the detailed telemetry—the accuracy scores, the loss curves, or whatever key variables were logged during that specific experiment runtime.

System Status & Permissions
Finally, before you start running complex queries, you should always check your credentials. Call get_user to verify your service account status and confirm that your agent has the proper access scope for all the projects it needs to talk to. It's a quick way to make sure everything's ready.

Here’s what you can do with this server: You can tell your AI client to list every project available, then grab detailed metadata on one of them using get_project. To find specific model outputs, it uses list_models to give you a packaged inventory. If the models are older, you'll use search_runs, which finds all tracked experiment runs in that project.

From those found runs, your agent can grab granular data points—like loss curves or final accuracy—by calling get_attributes. You never forget how to check permissions either; running get_user confirms the service account is active and ready for work.