mdl-train-mcp

An MCP server for monitoring and managing training jobs on Modal. Built for LLMs that need to check on long-running GPU training without drowning in log output.

Why?

Training logs on Modal can be tens of thousands of lines — weight loading bars, Omniverse init spam, 8 ranks of identical output. Dumping all of that into an LLM context is wasteful and often hits resource limits.

This server gives you browsable logs: start with a summary, then drill into what matters.

The get_logs workflow

Instead of returning a giant blob, get_logs has three modes:

1. Summary (default) — returns line count, first/last 10 lines, and any errors with line numbers. Small response, always works.

get_logs(app_id="ap-xxx")
→ {total_lines: 30000, errors: [{line: 847, text: "CUDA error: ..."}], head: [...], tail: [...]}

2. Window — read a specific range. Like scrolling through a file.

get_logs(app_id="ap-xxx", window_start=840, window_size=30)
→ 30 lines around the error

3. Grep — search with regex and context lines. Like grep -C.

get_logs(app_id="ap-xxx", grep="Error|Traceback", grep_context=15)
→ all errors with 15 lines of surrounding context

Landmarks — pass landmark_patterns in summary mode to get a table of contents:

get_logs(app_id="ap-xxx", landmark_patterns=["Iteration \\d+", "success_rate", "checkpoint"])
→ landmarks: [{line: 200, text: "Iteration 1/3000"}, {line: 5000, text: "success_rate: 0.95"}, ...]

Landmark sampling is fair across patterns — one pattern won't dominate.

Features

• Progress bar collapsing — tqdm bars, HF weight loading, and downloads are collapsed to their latest update (50 progress lines → 1 showing current state)
• Auto-retry on resource limits — if Modal's API rejects a large tail, automatically retries with smaller values and tells you what happened
• Error deduplication — 10,000 identical [Error] lines become a handful of unique entries
• Case-sensitive error detection — won't false-positive on metric names like rot_align_error

Setup

1. Install

# Using uv (recommended)
uv pip install mdl-train-mcp

# Or from source
git clone https://github.com/JoshuaSP/mdl-train-mcp
cd mdl-train-mcp
uv venv && uv pip install -e .

2. Configure Modal

Make sure you have the Modal CLI installed and authenticated:

pip install modal
modal setup

3. Add to Claude Code

Add to your .mcp.json:

{
  "mcpServers": {
    "mdl": {
      "command": "mdl-train-mcp",
      "env": {
        "MODAL_PROFILE": "your-profile"
      }
    }
  }
}

Or from source:

{
  "mcpServers": {
    "mdl": {
      "command": "uv",
      "args": ["--directory", "/path/to/mdl-train-mcp", "run", "mdl-train-mcp"],
      "env": {
        "MODAL_BIN": "/path/to/modal",
        "MODAL_PROFILE": "your-profile"
      }
    }
  }
}

Environment variables

Tools reference

list_apps

list_apps(state?: string, name_contains?: string)

Filter by state ("running", "deployed", "stopped", "ephemeral") or name substring.

get_logs

get_logs(
  app_id: string,
  tail?: number,              # log entries to fetch (default 500, max 5000)
  since?: string,             # "1h", "30m", "2d", or ISO datetime
  until?: string,
  source?: string,            # "stdout", "stderr", "system"
  window_start?: number,      # line number for window mode
  window_size?: number,       # lines to return (default 50, max 200)
  grep?: string,              # regex search (case-insensitive)
  grep_context?: number,      # context lines around matches (max 30)
  landmark_patterns?: string[] # regex patterns for summary landmarks
)

stop_app

stop_app(app_id: string)

Irreversible — terminates the app and all its containers.

License

MIT