New & Experimental - MCP support is a new feature. Please report any issues to support@trainy.ai or on Discord.

Model Context Protocol (MCP) Integration

The Pluto MCP server allows AI coding assistants to directly query your ML experiment data. This enables powerful workflows like:

Asking your AI assistant to analyze training runs and identify issues
Comparing metrics across experiments in natural language
Debugging failed runs by querying logs and metrics
Getting insights about your ML experiments without leaving your editor

Want to ask these questions inside Pluto instead of in your editor? Chat runs the same tools from the Pluto UI, driven by the Claude Code or Codex already on your machine — and it sets the MCP server up for you.

What is MCP?

The Model Context Protocol (MCP) is an open standard that enables AI assistants to securely access external data sources. With Pluto’s MCP server, tools like Claude Code can directly query your experiment data, metrics, and logs.

Setup with Claude Code

Prerequisites

A Pluto account with an API key (get one here)
Claude Code CLI installed

Step 1: Get Your API Key

Go to pluto.trainy.ai and sign in
Navigate to Settings > Developers > API Keys
Create a new API key and copy it

Step 2: Configure Claude Code

Run the following command to add the Pluto MCP server:

claude mcp add pluto --transport http --header "Authorization: Bearer mlps_xxxxxxxxxx" -- https://pluto-mcp.trainy.ai/mcp/

Replace mlps_xxxxxxxxxx with the API key you copied in Step 1.

Alternative: Manual Configuration

You can also manually edit the MCP settings file at ~/.claude/mcp_settings.json:

{
  "mcpServers": {
    "pluto": {
      "type": "http",
      "url": "https://pluto-mcp.trainy.ai/mcp/",
      "headers": {
        "Authorization": "Bearer your-api-key-here"
      }
    }
  }
}

If the file doesn’t exist, create it. Make sure the JSON is valid - you can verify with cat ~/.claude/mcp_settings.json | jq .

Step 3: Restart Claude Code

After saving the configuration, restart Claude Code for the changes to take effect:

claude

You should see the Pluto tools become available. You can verify by asking Claude: “What Pluto projects do I have?”

Available Tools

Once connected, the following tools become available to your AI assistant.

Projects, runs, and data

Tool	Description
`list_projects`	List all projects in your organization
`list_runs`	Search and filter experiment runs by project, name, or tags. Also accepts `config_filter` (see Filtering by config) and `konduktor_job_prefix` to find the run(s) for a Konduktor job — pass a full hashed job ID (e.g. `train-model-db87`, returns 0 or 1) or a YAML base name (e.g. `train-model`, returns the whole family). When set, `project_name` is optional and the search scopes to your org.
`get_run`	Get detailed information about a specific run. For runs launched on Konduktor, the output includes a labeled Konduktor Job block (job name and namespace, accelerator, nodes, GPUs/node, rank, node, restart attempt).
`list_metric_names`	List the distinct metric names logged in a project, with optional substring search
`query_metrics`	Query time-series metrics (loss, accuracy, etc.). Accepts `step_min` / `step_max` to scope to a step range
`query_logs`	Query console logs (stdout/stderr) from a run
`get_files`	Get files and artifacts from a run, with presigned download links
`get_statistics`	Per-metric count, min, max, mean, standard deviation, and last value

Comparing and ranking

Tool	Description
`compare_runs`	Compare a metric across multiple runs
`get_leaderboard`	Rank the runs in a project by a metric — aggregate by `MIN`, `MAX`, `AVG`, `LAST`, or `VARIANCE` (useful for ranking training stability), ascending or descending
`generate_comparison_url`	Build a shareable Pluto web-UI link with a set of runs pre-selected

Charts

Tool	Description
`visualize_metrics`	Render a line chart of one run’s metrics as a PNG, returned inline so the assistant can read it
`visualize_comparison`	Render a chart overlaying one metric across up to 20 runs

Editing runs

Tool	Description
`add_tags`	Add tags to a run
`remove_tags`	Remove tags from a run
`update_notes`	Set or clear a run’s notes

Dashboards

Tool	Description
`list_dashboards`	List a project’s saved dashboards, with a summary of each section
`get_dashboard`	Get one dashboard’s full layout: every section and widget
`read_dashboard`	Read a whole dashboard in one call. See Reading a dashboard
`render_panels`	Draw chosen panels from `read_dashboard` full-size, for a closer look before drawing conclusions from a specific value
`suggest_dashboard`	Propose a dashboard from the metrics a project actually logs: train/val pairs on shared charts, headline metrics as single-value tiles, a section per namespace, and system metrics collapsed at the bottom. Nothing is saved unless you pass `save_as`
`create_dashboard`	Create a dashboard from a short list of sections and metrics; widget IDs and grid layout are generated
`update_dashboard`	Rename a dashboard, replace its sections, or add sections, remove sections and add widgets without touching the rest
`list_dashboard_versions`	List a dashboard’s saved versions
`get_dashboard_version`	Get the layout of one past version
`restore_dashboard_version`	Save a past version as a new current version

Widgets the assistant creates can carry a description (up to 500 characters) of what the panel shows and what healthy looks like, for example “Should fall smoothly; a plateau means the learning rate is too low.” read_dashboard repeats it next to the panel, so a later reader gets the author’s intent as well as the curve.

The run-editing and dashboard tools modify your data. A connected assistant can change a run’s tags and notes, and create, edit or restore dashboards. It cannot create or delete runs, or alter their config, metrics, or files.Dashboard edits follow the same rules as the web UI: only a dashboard’s creator or an org admin or owner can change it, and only admins and owners can make one the project default. Every edit is a new version, so you can always restore the dashboard as it was before. An edit made against an out-of-date copy of the dashboard is rejected rather than overwriting someone else’s newer changes.Note that run groups are expressed as group: tags, so editing tags can move a run into or out of a group.

Reading a dashboard

read_dashboard takes a dashboard and a set of runs and reads every panel at once. Dynamic sections are expanded against the project’s real metric names. Reading 50 per-layer gradient-norm panels takes one call rather than 50. Choose how much to return with detail:

`detail`	Returns
`index`	Panel titles, descriptions and links. No data is fetched
`digest` (default)	Adds a one-line summary of each curve: trend, spikes located by step, plateaus, non-finite values
`full`	Adds contact-sheet images with up to 20 panels each

Every response states its coverage (for example panels 1-20 of 47), so a partial read is never mistaken for the whole dashboard. A panel that can’t be read, such as one whose metric was renamed, reports the problem at its own position without affecting the others. Each panel prints a key= token that render_panels accepts, and it keeps pointing at the same panel after the dashboard is edited. Python panels come back as their source code, never as an image.

“Read the Optimizer Debug dashboard for runs MCR-1 and MCR-2 — do any layers’ gradient norms look wrong?”

Filtering by config and system metadata

list_runs accepts a config_filter that filters server-side on any field in a run’s config or captured environment. The simplest form is a plain string, <key> <operator> <value>:

trainer.lr > 0.001
model.name is dit
checkpoint.r2_prefix contains 37a
systemMetadata.gpu_model is H100

Keys are dot-paths into the run’s config by default. Prefix a key with systemMetadata. to target the captured environment (GPU model, CUDA version, git branch, hostname, …) instead. Supported operators: contains, does not contain, is, is not, starts with, ends with, regex, >, <, >=, <=. Pass a list of conditions to require all of them — conditions always AND together, there is no OR:

["model.name is dit", "trainer.lr > 0.001"]

config_filter reaches config.* and systemMetadata.* only. To filter on metric values (summaryMetrics.*), use the filters query language instead.

Querying a step range

query_metrics and visualize_metrics accept step_min and step_max (inclusive) to scope a query to a window of steps — handy for zeroing in on a loss spike or an anomaly rather than pulling the whole series:

“Plot train/loss between steps 4000 and 4500 for run MMP-42”

Example Prompts

Once configured, you can interact with your Pluto data naturally through your AI assistant: List your projects:

“What ML projects do I have in Pluto?”

Find recent runs:

“Show me the last 5 training runs in the gpt-finetuning project”

Analyze a run:

“What was the final loss for run 1234? Did it converge?”

Compare experiments:

“Compare the train/loss between runs 100, 101, and 102 - which performed best?”

Debug failures:

“Show me the error logs from run 456 - why did it fail?”

Build a dashboard:

“Suggest a dashboard for my-project, then save it as Training Overview”

Get insights:

“Are there any anomalies in the metrics for my latest run?”

Claude Code Web

To make the pluto MCP available to your Claude Code web UI and cloud agents, visit: https://claude.ai/settings/connectors?modal=add-custom-connector From here you can enter the following information for the Pluto MCP server:

Name: Pluto
Remote MCP Server URL: https://pluto-mcp.trainy.ai/mcp

Afterwards, click on Connect and you will be redirected to an OAuth flow where you can enter the Pluto API key you want to use for the MCP connection.

Feedback

MCP integration is experimental. We’d love to hear your feedback:

Discord: Join our community
Email: support@trainy.ai

Get Started

Core

API Reference

Integrations

Migrate

Changelog

MCP Integration

Model Context Protocol (MCP) Integration

What is MCP?

Setup with Claude Code

Prerequisites

Step 1: Get Your API Key

Step 2: Configure Claude Code

Step 3: Restart Claude Code

Available Tools

Projects, runs, and data

Comparing and ranking

Charts

Editing runs

Dashboards

Reading a dashboard

Filtering by config and system metadata

Querying a step range

Example Prompts

Claude Code Web

Feedback

​Model Context Protocol (MCP) Integration

​What is MCP?

​Setup with Claude Code

​Prerequisites

​Step 1: Get Your API Key

​Step 2: Configure Claude Code

​Step 3: Restart Claude Code

​Available Tools

​Projects, runs, and data

​Comparing and ranking

​Charts

​Editing runs

​Dashboards

​Reading a dashboard

​Filtering by config and system metadata

​Querying a step range

​Example Prompts

​Claude Code Web

​Feedback

Model Context Protocol (MCP) Integration

What is MCP?

Setup with Claude Code

Prerequisites

Step 1: Get Your API Key

Step 2: Configure Claude Code

Step 3: Restart Claude Code

Available Tools

Projects, runs, and data

Comparing and ranking

Charts

Editing runs

Dashboards

Reading a dashboard

Filtering by config and system metadata

Querying a step range

Example Prompts

Claude Code Web

Feedback