Which integrations are available for this server?

Allows locating and exporting .onnx model files and checkpoints from trained ML-Agents models. Allows controlling Unity ML-Agents training runs: launch, stop, resume, monitor progress, compare runs, and manage configurations through natural language.

How do I use mlagents-mcp?

1. Click on "Install Server". 2. Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state. 3. In the chat, type @ followed by the MCP server name and your instructions, e.g., "@mlagents-mcp Start a new training run with default settings." That's it! The server will respond to your query, and you can continue using it as needed. Here is a step-by-step guide with screenshots.

mlagents-mcp

by limam-B

Overview Schema Related Servers Score Discussions

Python

Local

mlagents-mcp

MCP server for controlling Unity ML-Agents training runs from Claude Code.

Launch, stop, resume, monitor, compare, and export ML-Agents training — all through natural conversation without leaving your editor.

Features

Training control — start, stop, resume runs as background processes
Instant checks — check training progress without blocking the conversation
Live monitoring — read TensorBoard metrics, reward curves, and training logs in real time
Run comparison — compare metrics across runs for hyperparameter tuning
Config management — read and deep-merge update YAML training configs
Model export — locate .onnx models and checkpoints
Two training modes — Unity Editor (interactive) and built executable (headless batch)

Related MCP server: claude-sessions-mcp

Requirements

Python 3.10+
uv package manager
Unity ML-Agents mlagents-learn available in PATH (or via conda env)
Claude Code CLI

Installation

git clone https://github.com/limam-B/mlagents-mcp-server.git
cd mlagents-mcp-server
uv sync

Quick setup with Claude Code

# Register the MCP server:
claude mcp add --scope project --transport stdio mlagents-training \
  -- uv run --directory /path/to/mlagents-mcp-server mlagents-mcp

# Unregister (from all scopes to clean up stale configs):
claude mcp remove --scope local mlagents-training
claude mcp remove --scope user mlagents-training
claude mcp remove --scope project mlagents-training

# List registered servers:
claude mcp list

With environment variables

The server reads its configuration from environment variables. Add them to your .mcp.json (project-level) or pass them via the CLI:

{
  "mcpServers": {
    "mlagents-training": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/mlagents-mcp-server", "mlagents-mcp"],
      "env": {
        "MLAGENTS_PROJECT_ROOT": "/path/to/your/unity/project",
        "MLAGENTS_RESULTS_DIR": "results",
        "MLAGENTS_CONFIG_DIR": "config"
      }
    }
  }
}

With conda (if ML-Agents is installed in a conda env)

{
  "mcpServers": {
    "mlagents-training": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/mlagents-mcp-server", "mlagents-mcp"],
      "env": {
        "MLAGENTS_PROJECT_ROOT": "/path/to/your/unity/project",
        "MLAGENTS_RESULTS_DIR": "results",
        "MLAGENTS_CONFIG_DIR": "config",
        "MLAGENTS_CONDA_ENV": "mlagents",
        "MLAGENTS_CONDA_PATH": "/home/user/miniconda3"
      }
    }
  }
}

Environment variables

Variable	Default	Description
`MLAGENTS_PROJECT_ROOT`	`.`	Root directory of your Unity project
`MLAGENTS_RESULTS_DIR`	`results`	Training results directory (relative to project root)
`MLAGENTS_CONFIG_DIR`	`config`	Training config YAML directory (relative to project root)
`MLAGENTS_CONDA_ENV`	—	Conda environment name to activate before running `mlagents-learn`
`MLAGENTS_CONDA_PATH`	—	Path to conda installation (e.g. `/home/user/miniconda3`)

Tools (18)

Training control

Tool	Description
`force_training`	Launch a new training run (overwrites previous results). Blocks until ready by default.
`stop_training`	Gracefully stop a run (SIGINT, saves the model).
`resume_training`	Resume from checkpoint. Auto-reads config from previous run.
`cleanup_processes`	Kill orphaned mlagents-learn and Unity build processes not tracked by any active run.

Monitoring

Tool	Description
`get_run_status`	Status overview: reward trend, checkpoints, step progress.
`get_metrics`	Read TensorBoard scalars (reward, losses, learning rate, etc.).
`get_training_logs`	Tail live stdout/stderr from an active run.
`list_runs`	List all known runs with status filtering.

Comparison & export

Tool	Description
`compare_runs`	Compare a metric across multiple runs (min/max/final + trend).
`export_model`	Locate .onnx model files and checkpoints.

Configuration

Tool	Description
`get_config`	Read a YAML training config.
`update_config`	Deep-merge updates into a config (only specified keys change).

Wait & check

Tool	Default	Description
`wait_for_first_metrics`	Blocks	Blocks until first TensorBoard data point appears (~1-2 min).
`wait_for_completion`	Blocks	Blocks until training finishes. For automated run chaining (up to 4 hours).
`check_step`	Instant	Check if training reached a target step. Returns current progress.
`check_reward`	Instant	Check if mean reward reached a target. Returns current reward.
`check_completion`	Instant	Check if training finished. Returns current status and progress.
`check_checkpoint`	Instant	Check if new .onnx files appeared. Returns checkpoint list.

The check_* tools always return instantly — they never block the conversation. Use wait_for_completion when you want to block until a run finishes (e.g. to chain skill A → skill B automatically).

Two training modes

Editor mode (no `env_path`)

Training connects to the Unity Editor. force_training blocks until mlagents-learn prints "Listening on port... press Play", then you (or an AI agent) presses Play in Unity.

force_training(config_path="movement.yaml", run_id="Movement_v1")
# → blocks until "Listening on port 5004. Start training by pressing Play..."

Batch mode (with `env_path`)

Training launches a built executable directly — no Unity Editor needed. force_training blocks until the executable connects.

force_training(
    config_path="movement.yaml",
    run_id="Movement_v1",
    env_path="/path/to/Build.x86_64",
    num_envs=12,
    no_graphics=True,
)
# → blocks until "Connected to Unity environment"

Example workflow

A typical automated training session:

1. force_training(config, run_id, ...)     # launch, blocks until ready
2. wait_for_first_metrics(run_id)          # blocks until data flowing
3. wait_for_completion(run_id)             # blocks until training ends (hours)
4. export_model(run_id)                    # get checkpoint
5. update_config(next_skill, init_path=..) # chain checkpoint to next skill
6. [repeat from step 1 for next skill]

Steps 1-2 block briefly during startup. Step 3 blocks for the full duration (hours) — use this for automated chaining. For manual monitoring, use check_step/check_completion instead of step 3.

Development

# Install dev dependencies:
uv sync --group dev

# Lint:
uv run ruff check src/

# Format:
uv run ruff format src/

# Run the server directly (stdio):
uv run mlagents-mcp

Project structure

src/mlagents_mcp/
  server.py            # FastMCP app, all 18 tool definitions, entry point
  process_manager.py   # Subprocess launch/stop, log capture, port assignment
  metrics_reader.py    # TensorBoard event file parsing
  config_manager.py    # YAML config read/write/deep-merge
  run_registry.py      # Thread-safe run tracking + historical disk scan
  waiters.py           # Blocking wait logic for all wait_for_* tools
  types.py             # Shared dataclasses and enums

License

MIT

This server cannot be installed

license - not found

quality - not tested

maintenance

How are these scores calculated?

Maintenance

–Maintainers

–Response time

–Release cycle

–Releases (12mo)

Commit activity

Resources

GitHub Repository

Need Help?

Related Servers

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Latest Blog Posts

Your AI Chatbot Just Exposed Your CEO's Salary to an Intern
By Om-Shree-0709 on July 2, 2026.
Agent Identity
MCP Security
OAuth Delegation
Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)
By Om-Shree-0709 on June 30, 2026.
Agentic Ai
Prompt Injection
WebAssembly
Lightport: Open-Sourcing Glama's AI Gateway
By punkpeye on April 27, 2026.
OpenAI
open source

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/limam-B/mlagents-mcp-server'

If you have feedback or need assistance with the MCP directory API, please join our Discord server

mlagents-mcp

Features

Requirements

Installation

Quick setup with Claude Code

With environment variables

With conda (if ML-Agents is installed in a conda env)

Environment variables

Tools (18)

Training control

Monitoring

Comparison & export

Configuration

Wait & check

Two training modes

Editor mode (no env_path)

Batch mode (with env_path)

Example workflow

Development

Project structure

License

Maintenance

Resources

Looking for Admin?

Latest Blog Posts

MCP directory API

Editor mode (no `env_path`)

Batch mode (with `env_path`)