Control NVIDIA Isaac Sim robotics simulator through MCP with 42 tools
NVIDIA Isaac Sim MCP Server
This MCP server, io.github.whats2000/isaacsim-mcp-server, controls the NVIDIA Isaac Sim robotics simulator via the Model Context Protocol. It provides 42 tools for interacting with Isaac Sim, and is distributed as a Python package with a described compatibility of Python 3.10+ and Isaac Sim 5.1.0 through 6.0.1.
๐ ๏ธ Key Features
MCP-based control of NVIDIA Isaac Sim
42 available tools
Tooling distributed as isaacsim-mcp-server (PyPI)
๐ Use Cases
Programmatic robotics simulation control through MCP
Integration of Isaac Sim into MCP-driven developer workflows
โก Developer Benefits
Standardized interface via Model Context Protocol
Python 3.10+ compatibility
MIT-licensed project
โ ๏ธ Limitations
Isaac Sim version scope is indicated as 5.1.0โ6.0.1
Connect any MCP-compatible IDE (Cursor, VS Code, Claude Code, Windsurf, Antigravity, JetBrains) to a running Isaac Sim instance and control it with plain-English prompts -- create robots, build scenes, run simulations, and debug physics all from your editor.
107+ robots auto-discovered from the Isaac Sim asset library (Franka, UR, Unitree, Boston Dynamics, and more)
Step-and-observe debugging -- step the simulation and inspect prim positions, joint states, and physics in one call
Hot-reload -- iterate on Python controllers without restarting Isaac Sim
Multi-instance -- run multiple Isaac Sim sessions side by side on different ports
Built for Isaac Sim 5.1.0 - 6.0.1 (PhysX + Newton) with a modular adapter layer for version isolation
Installation
Option A: pip install (recommended)
bash
pip install isaacsim-mcp-server
This installs the MCP server and the isaacsim-mcp-server CLI. You still need the Isaac Sim extension from the repo (see Launching Isaac Sim below).
Option B: From source
bash
git clone https://github.com/whats2000/isaacsim-mcp-server
cd isaacsim-mcp-server
./scripts/setup_python_env.sh
Requirements
Requirement
Version
NVIDIA Isaac Sim
5.1.0 - 6.0.1 (PhysX or Newton)
Python
3.10+
uv
latest (for source install)
Platform
Linux (Ubuntu 22.04+) or Windows 10/11
IMPORTANT
Linux and Windows are supported. On Windows, use the PowerShell
launcher scripts/run_isaac_sim.ps1 in place of the .sh scripts (see below).
macOS is not supported because NVIDIA Isaac Sim does not run on macOS.
NOTE
We are welcoming contributions to support other Isaac Sim versions.
The adapter layer is designed for easy version isolation.
Quick Start
1. Set up the environment
If you installed from source:
bash
./scripts/setup_python_env.sh
On Windows, uv sync creates the virtual environment (.venv) and installs
the package plus its dependencies:
powershell
uv sync
2. Launch Isaac Sim with the extension
bash
./scripts/run_isaac_sim.sh
You should see in the logs:
code
Registered 42 command handlers
Isaac Sim MCP server started on localhost:8766
The script looks for Isaac Sim in $HOME/isaacsim; set ISAACSIM_ROOT to use a
different install.
Choosing the physics engine. Isaac Sim 6.0+ ships PhysX (default) and Newton
backends. Select one with --newton / --physx, or ISAACSIM_ENGINE:
bash
./scripts/run_isaac_sim.sh # PhysX (default)
./scripts/run_isaac_sim.sh --newton # Newton
ISAACSIM_ENGINE=newton ./scripts/run_isaac_sim.sh
The same flags work with scripts/launch_isaac_sim_mcp.sh. Everything else on
the command line is forwarded to Kit untouched. The server auto-detects the
active engine, so no MCP-side configuration changes. Newton requires 6.0 or
newer; asking for it on 5.1.0 fails with a clear message.
On Windows, use the PowerShell launcher instead. It takes the same engine
selection and forwards extra arguments to Kit:
powershell
.\scripts\run_isaac_sim.ps1 # PhysX (default)
.\scripts\run_isaac_sim.ps1 -Engine newton # Newton
$env:ISAACSIM_ENGINE = 'newton'; .\scripts\run_isaac_sim.ps1
The script resolves the install from -IsaacSimRoot, then $env:ISAACSIM_ROOT,
then a local source build, then C:\isaacsim, then %USERPROFILE%\isaacsim. It
also creates a writable USD working directory (.cache\usd) since Windows has no
/tmp.
Optional: Beaver3D / NVIDIA API keys for 3D generation
bash
export BEAVER3D_MODEL="<your beaver3d model name>"export ARK_API_KEY="<your beaver3d api key>"export NVIDIA_API_KEY="<your nvidia api key>"
On Windows (PowerShell):
powershell
$env:BEAVER3D_MODEL = "<your beaver3d model name>"
$env:ARK_API_KEY = "<your beaver3d api key>"
$env:NVIDIA_API_KEY = "<your nvidia api key>"
3. Connect your IDE
Add the MCP server to your editor. Replace the path with your actual repo location.
The command examples are for Linux/macOS; each guide shows the Windows
equivalent, which wraps the PowerShell launcher scripts\run_mcp_server.ps1.
Claude Code (CLI)
bash
claude mcp add isaac-sim /path/to/isaacsim-mcp-server/scripts/run_mcp_server.sh
Open the agent side panel, click โฆ > MCP Servers > Manage MCP Servers >
View raw config, or edit ~/.gemini/config/mcp_config.json (global) or
.agents/mcp_config.json (workspace):
Go to Settings > Tools > AI Assistant > MCP Servers and add the server, with
the command /path/to/isaacsim-mcp-server/scripts/run_mcp_server.sh. See the
JetBrains MCP docs for details.
On Windows, set the command to powershell and the arguments to
-NoProfile -ExecutionPolicy Bypass -File C:\path\to\isaacsim-mcp-server\scripts\run_mcp_server.ps1.
4. Start prompting
text
Check the connection with get_scene_info.
If the scene is empty, create a physics scene.
Add a Franka robot at the origin and a Go1 quadruped at [2, 0, 0].
Architecture
text
MCP Client (IDE)
|
v
isaacsim-mcp-server (PyPI package / CLI)
|
v TCP socket (localhost:8766)
|
isaac.sim.mcp_extension (Omniverse extension)
|
v
Handlers -> Adapter -> Isaac Sim 5.1 / 6.0 APIs
Open defects a normal session can hit. Each is warned about at the
point of use where that is possible; this list is for choosing a runtime before
you start.
Affects
What happens
Issue
6.0 Newton
Joint drives do not converge โ a commanded target is overshot and the joint keeps going, and joint limits are not enforced. Scene setup, stepping and inspection are fine; run motion work on PhysX (isaac-sim.sh).
get_lidar_point_cloud fills on roughly a third of reads, so a caller must retry. A lidar created while the timeline is running never fills at all โ create it stopped.
Check the connection with get_scene_info. If the scene is empty, create a physics scene.
Add stronger lighting and place a camera that looks at the workspace.
Robot layout
text
Create three Franka robots in a row at [0,0,0], [2,0,0], and [4,0,0].
Then add a Go1 robot at [1, 3, 0].
Environment loading
text
List available environments, choose a warehouse-like one, and load it.
Create a camera and capture an image.
Asset search and 3D generation
text
Search for a rusty desk, load the best result near [0, 5, 0], scaled to [2, 2, 2].
Advanced Usage
Multiple Instances
Run multiple Isaac Sim sessions side by side. Each uses a different port (auto-assigned from 8766).
bash
# First instance (default port 8766)
claude mcp add isaac-sim /path/to/isaacsim-mcp-server/scripts/run_mcp_server.sh
# Second instance (port 8767)
claude mcp add isaac-sim-2 -e ISAAC_MCP_PORT=8767 -- /path/to/isaacsim-mcp-server/scripts/run_mcp_server.sh
Install a dedicated Isaac Sim MCP application icon:
bash
./scripts/install_desktop_entry.sh
This creates a launcher that auto-assigns ports, waits for the extension socket, and cleans up on exit.
Recommended Workflow
Start with get_scene_info to verify the connection
Create a physics scene if the stage is empty
Prefer purpose-built tools before execute_script
Use list_available_robots / list_environments before loading
Use create_action_graph to wire OnPlaybackTick โ ScriptNode controllers
Use step_simulation with observe_prims and observe_joints for debugging
Use reload_script to iterate on controllers without restarting
Demo: Franka Pick-and-Place
A ready-to-run demo at demo/franka_pick_place.py using RMPflow for motion planning:
text
Please use the Isaac MCP tool complete this:
Create a physics scene with a ground plane, then spawn a Franka FR3 robot at the origin.
Add two textured tables with a gap along Y. Place a small textured cube with physics enabled on top of the first table.
Use `create_action_graph` to wire `OnPlaybackTick` โ `ScriptNode`, and write a pick-and-place controller script using RMPflow for motion planning. Save the script to the `demo/` directory.
Use `get_prim_info` to query actual positions and sizes of the tables and cube before writing the controller โ do not hardcode coordinates.
Start the simulation with Play. The robot should pick the cube from table 1 and place it on table 2. Verify the process using `step_simulation` with `observe_prims` on the cube to confirm it reaches table 2.
Uses create_action_graph with script_file for one-step Action Graph + ScriptNode setup, plus the observability tools: get_joint_config, step_simulation with observe_prims, get_physics_state, and edit_action_graph for script hot-reload.
Development
bash
# Run the MCP inspector
./.venv/bin/python -m mcp dev ./isaac_mcp/server.py
The inspector is available at http://localhost:5173.
Engine selection lives in scripts/lib/isaac_launcher.sh: each engine maps to
the launcher script Isaac Sim ships for it. Adding an entry to that map is all a
new backend needs โ it enables both ISAACSIM_ENGINE=<name> and --<name> in
every launcher script.
Troubleshooting
If Isaac Sim says Can't find extension with name: isaac.sim.mcp_extension:
bash
# Make sure you're in the repo rootpwdtest -f ./isaac.sim.mcp_extension/config/extension.toml && echo OK
Note: --ext-folder must point to the repo root, not to isaac.sim.mcp_extension/ directly.
Contributing
Pull requests are welcome. Improvements to tools, docs, adapters, and tests are all useful.
License
MIT License. Copyright (c) 2023-2025 omni-mcp, Copyright (c) 2026 whats2000. See LICENSE.