Isaaclab

MCP server for NVIDIA Isaac Lab — RL training, environment management, and policy evaluation from Claude.

Runs locally on your Mac and communicates with Isaac Lab on a remote Brev GPU instance through an SSH tunnel. All heavy simulation stays on the GPU; Claude just sends commands.

Not Isaac Sim. This server controls Isaac Lab (RL environments, training pipelines, policy evaluation). For low-level Isaac Sim control (USD prims, scene authoring, Kit commands), see mcp-server-isaacsim.

Architecture

┌──────────┐    stdio    ┌──────────────┐   SSH tunnel   ┌─────────────────┐
│  Claude   │◄──────────►│  MCP Server  │◄──────────────►│  Remote Agent   │
│  (local)  │            │  (local Mac) │   port 8421    │  (Brev GPU)     │
└──────────┘            └──────────────┘                └────────┬────────┘
                                                                  │
                                                          ┌───────▼────────┐
                                                          │   Isaac Lab    │
                                                          │  (Isaac Sim)   │
                                                          └────────────────┘

MCP Server (this repo) runs on your Mac as a stdio MCP server. It opens an SSH tunnel to the Brev instance and forwards all requests to the Remote Agent — a FastAPI service running next to Isaac Lab on the GPU box.

Prerequisites

• Python 3.10+
• A Brev GPU instance (provision with brev create)
• SSH access to the instance (brev ssh)
• Isaac Lab installed on the instance (setup script included)

Quick Start

1. Install locally

git clone [email protected]:chloepilonv/mcp-server-isaaclab.git
cd mcp-server-isaaclab
pip install -e .

2. Provision a Brev GPU instance

brev create isaaclab-gpu --gpu A100
brev ssh isaaclab-gpu

3. Install Isaac Lab on the instance

# From your Mac:
scp scripts/setup-brev-isaaclab.sh ubuntu@<BREV_HOST>:~
ssh ubuntu@<BREV_HOST> bash ~/setup-brev-isaaclab.sh

This installs Isaac Lab + the skrl, rsl_rl, and sb3 RL frameworks.

4. Deploy the remote agent

./scripts/deploy-remote-agent.sh <BREV_HOST> ubuntu ~/.ssh/your_key

This copies the agent code, installs it, and starts it as a systemd service on port 8421.

5. Configure Claude

The project includes .mcp.json so Claude Code automatically picks up the server when you're in this directory.

For Claude Desktop, add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "isaaclab": {
      "command": "mcp-server-isaaclab"
    }
  }
}

For Claude Code in other projects, add to the project's .mcp.json:

{
  "mcpServers": {
    "isaaclab": {
      "command": "mcp-server-isaaclab"
    }
  }
}

Tools

Connection

Tool	Description
connect_instance	Establish SSH tunnel to Brev GPU instance
disconnect_instance	Tear down the connection
instance_status	GPU utilization, active sessions & jobs
gpu_status	Detailed GPU memory, temperature, utilization

Simulation (Interactive)

Tool	Description
list_environments	List all registered Isaac Lab tasks
create_session	Create an interactive simulation session
step_session	Step simulation forward (random or specified actions)
reset_session	Reset environment to initial state
get_observation	Get current observations + action/obs space info
close_session	Close session and free GPU memory

Training

Tool	Description
start_training	Launch an async RL training job
monitor_training	Get status, recent logs, latest checkpoint
get_training_logs	Read full training logs
stop_training	Stop a running training job
list_training_jobs	List all jobs (running, completed, failed)

Evaluation & Files

Tool	Description
evaluate_policy	Evaluate a checkpoint, optionally record video
list_checkpoints	Browse saved model checkpoints
list_log_dirs	Browse training log directories
list_videos	List recorded simulation videos
read_remote_file	Read any text/image file on the instance
run_isaaclab_script	Run arbitrary Isaac Lab Python scripts

Example Conversations

Train a locomotion policy:

> Connect to my Brev instance at 203.0.113.42
> What environments are available for quadruped locomotion?
> Train Anymal-D on rough terrain with rsl_rl, 4096 envs, 1500 iterations
> Check on the training
> Evaluate the best checkpoint and record a video

Explore an environment interactively:

> Connect to my Brev GPU
> Create a session with Isaac-Cartpole-v0, 32 envs
> What does the observation space look like?
> Step 100 times with random actions — what are the rewards?
> Reset and try again
> Close the session

Monitor GPU and manage jobs:

> What's the GPU status?
> List all training jobs
> Stop the Ant training — it's not converging
> Show me the last 200 lines of logs from the Franka training

Supported RL Frameworks

Framework	Best For	Notes
skrl	General purpose	Modern, modular, good default choice
rsl_rl	Locomotion	ETH RSL's framework, optimized for legged robots
sb3	Prototyping	Stable Baselines 3, easy to use
rl_games	Multi-GPU	NVIDIA's framework, scales well

Available Environments (selection)

Category	Examples
Classic	Isaac-Cartpole-v0, Isaac-Ant-v0, Isaac-Humanoid-v0
Manipulation	Isaac-Reach-Franka-v0, Isaac-Lift-Cube-Franka-v0, Isaac-Open-Drawer-Franka-v0
Locomotion	Isaac-Velocity-Flat-Anymal-D-v0, Isaac-Velocity-Rough-Unitree-Go2-v0
Navigation	Isaac-Navigation-Flat-Anymal-C-v0

Use list_environments to get the full list from your installation.

Project Structure

mcp-server-isaaclab/
├── src/mcp_server_isaaclab/
│   ├── server.py            # MCP server (runs locally, exposes tools)
│   ├── connection.py        # SSH tunnel + HTTP client manager
│   └── remote/
│       └── agent.py         # FastAPI agent (runs on Brev GPU)
├── scripts/
│   ├── deploy-remote-agent.sh      # Deploy agent to Brev
│   └── setup-brev-isaaclab.sh      # Install Isaac Lab on instance
├── .mcp.json                # Claude Code MCP config
├── pyproject.toml
└── README.md

Development

pip install -e ".[dev]"
ruff check src/
pytest

License

MIT