roz uses the Model Context Protocol (MCP) to discover and call robot-specific tools at runtime. Each Docker sim container bundles an MCP server on port 8090 that exposes tools tailored to the robot’s capabilities.
When you start a simulation with roz sim start, the following happens:
- The Docker container starts with Gazebo, the robot’s middleware stack, and an MCP server.
- The agent connects to the MCP server on port 8090.
- The agent calls
tools/list to discover available tools and their schemas.
- Discovered tools are registered with the agent session via
RegisterTools.
Tools are scoped to the container that provides them. When a container stops, its tools are automatically unregistered from the session.
ros2-manipulator
Tools for controlling a 6-DOF arm (UR5) via MoveIt2.
| Tool | Description | Parameters |
|---|
get_joint_state | Returns current joint positions and velocities for all 6 joints. | None |
move_to_named_target | Moves the arm to a predefined named pose (e.g. "home", "up", "ready"). | target: string |
move_to_pose | Plans and executes a motion to a Cartesian pose. | x, y, z: float, qx, qy, qz, qw: float |
stop_arm | Immediately stops all arm motion. | None |
# Example: move to home position
# The agent calls this tool automatically when you say "move to home"
{
"name": "move_to_named_target",
"parameters": { "target": "home" }
}
px4-gazebo-humble
The PX4 drone container uses gRPC (MAVSDK) for flight commands rather than MCP. The agent communicates with the PX4 autopilot through a gRPC bridge that translates high-level commands to MAVLink.
| Command | Description | Parameters |
|---|
arm | Arms the drone motors. | None |
takeoff | Arms and takes off to a specified altitude. | altitude: float |
land | Lands at current position. | None |
go_to | Flies to a GPS or local coordinate. | lat, lon, alt: float |
set_offboard | Enters offboard mode for WASM velocity control. | None |
PX4 flight commands go through gRPC, not MCP. The agent handles this difference transparently — you give the same natural language instructions regardless of the underlying protocol.
ardupilot-gazebo
The ArduPilot drone container provides similar flight commands through its own interface.
| Command | Description | Parameters |
|---|
arm | Arms the drone motors. | None |
takeoff | Arms and takes off to a specified altitude. | altitude: float |
land | Lands at current position. | None |
go_to | Flies to a local coordinate. | x, y, z: float |
ros2-nav2
Tools for controlling a differential drive mobile robot via the Nav2 stack.
| Tool | Description | Parameters |
|---|
navigate_to | Sends a navigation goal to Nav2. The robot plans a path and drives to the target. | x, y: float, yaw: float |
follow_waypoints | Sends a sequence of waypoints for the robot to follow in order. | waypoints: [{x, y, yaw}] |
Custom MCP Servers
If you build a custom sim container, you can bundle your own MCP server. The agent discovers tools from any MCP server that:
- Listens on port 8090 inside the container.
- Implements the MCP
tools/list and tools/call endpoints.
- Returns valid JSON Schema for each tool’s parameters.
The MCP server can also provide a substrate:workflow prompt that gives the agent context about how to use the tools together. This prompt is injected into the agent’s system context when the tools are registered.
MCP vs WASM
MCP tools operate at 1-3Hz — they are discrete actions like “move to this pose” or “navigate to this point.” For continuous, closed-loop control at 100Hz, the agent writes WASM controllers that use the channel interface instead.
The agent chooses between MCP tools and WASM controllers based on the task. Simple discrete actions use MCP. Continuous or reactive behaviors use WASM. The agent can combine both in a single session. 
