webots-mcp-kit 0.9.15

Windows-first Webots developer toolkit with CLI, MCP server, controller SDK, and bundled benchmarks.

webots-mcp-kit

webots-mcp-kit is a Windows-first developer toolkit for connecting LLM agents to Webots.

It combines:

• a pip-installable CLI
• an MCP server with Webots session tools
• a controller-side SDK for structured telemetry and manual overrides
• bundled example scenarios and benchmarks

The bundled line follower is not the product itself. It is the first reference example. The toolkit is meant to be reusable across other Webots robots, controllers, and worlds.

Current release

v0.9.15

Current focus:

• Windows-first local development
• Webots R2025a
• direct Webots integration without ROS2
• stable controller-side integration through ControllerAgent
• bundled scenarios plus registry-backed benchmark thresholds
• hosted-safe CI with separate self-hosted runtime smoke
• release pipeline for GitHub Release, TestPyPI, and PyPI
• public-contract regression coverage and external-user onboarding
• MCP reliability for LLM-driven Webots sessions

Current support

Supported now:

• Windows
• Webots R2025a
• Python 3.11+
• interactive self-hosted runtime through the interactive-webots runner label

Not supported in the pre-v1.0.0 line:

• Windows service runner runtime
• Linux/macOS runtime support
• ROS2 integration
• multi-robot orchestration

Bundled scenarios

• line-follower
  • camera-based line tracking
  • example world: examples/line-follower
• obstacle-avoidance
  • proximity-sensor obstacle avoidance
  • example world: examples/obstacle-avoidance
• waypoint-nav
  • fixed-waypoint navigation in an open arena
  • example world: examples/waypoint-nav

Install

Use an isolated virtual environment. The toolkit depends on mcp, which may pull shared web stack packages into your global Python install.

python -m venv .venv
.venv\Scripts\Activate.ps1
pip install webots-mcp-kit

Development install:

python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -e .[dev]

Optional CI or slow-machine tuning:

$env:WEBOTS_KIT_SESSION_START_TIMEOUT='90'

Quick start

I want to connect an agent to Webots

webots-kit doctor
webots-kit session start --scenario line-follower --controller example --mode fast --render off
webots-kit mcp serve
webots-kit benchmark list
webots-kit benchmark run line-follower --controller example --output .\report.json --duration-s 3
webots-kit benchmark report .\report.json

I want to integrate my own controller

webots-kit controller scaffold .\controllers\my_agent.py --scenario line-follower
webots-kit controller validate .\controllers\my_agent.py --scenario line-follower --strict --json
webots-kit benchmark run line-follower --controller .\controllers\my_agent.py --output .\report.json

CLI surface

• webots-kit doctor [--json]
• webots-kit session start --scenario <name> --world <path> --controller <path-or-id> [--robot-name <name>] [--robot-def <def>] [--mode fast|realtime|pause] [--render on|off]
• webots-kit session inspect --session <id>
• webots-kit session logs --session <id> [--name <file>] [--tail <n>]
• webots-kit session stop --session <id>
• webots-kit benchmark list
• webots-kit benchmark run <scenario> --controller <path-or-id> --output <report.json> [--duration-s <seconds>]
• webots-kit benchmark report <report.json>
• webots-kit controller validate <path> [--scenario <name>] [--strict] [--json]
• webots-kit controller scaffold <path> [--scenario <name>] [--force]
• webots-kit mcp serve

Controller integration

There are two supported usage paths:

• bundled example controllers under examples/
• user-supplied Python controllers that integrate with ControllerAgent

The public controller-side entrypoint is ControllerAgent. The stable public contract is:

• ControllerAgent.from_robot(...)
• begin_step()
• report_step(...)

Minimal integration shape:

from controller import Robot
from webots_mcp_kit.agent import ControllerAgent

robot = Robot()
agent = ControllerAgent.from_robot(robot, default_camera="camera")

while robot.step(int(robot.getBasicTimeStep())) != -1:
    override = agent.begin_step()
    # apply your control logic, optionally overriding wheel commands
    agent.report_step(
        sensors={},
        metrics={},
        actuators={},
        camera_frames=None,
    )

Use webots-kit controller validate <path> to check whether a controller follows the expected integration pattern. Use webots-kit controller scaffold when you want a working starter file based on a bundled scenario.

MCP tools

• webots_session_start
• webots_session_stop
• webots_list_robots
• webots_list_devices
• webots_get_state
• webots_get_sensors
• webots_capture_camera
• webots_set_motor_velocity
• webots_step
• webots_pause_resume
• webots_reset
• webots_run_benchmark

Stable payload shapes:

• webots_list_devices -> { robot, scenario, devices }
• webots_get_sensors -> { robot, scenario, state, sensors, metrics, actuators, meta }
• failed MCP tool calls -> { ok: false, error: { code, message, details, retriable } }

Reference docs:

• First hour guide
• MCP contracts
• Self-hosted runtime smoke
• Custom controller integration
• Add a new scenario
• Failed runtime smoke triage
• PyPI install and upgrade
• Release checklist
• Runner maintenance
• Packaging verification

Testing

Unit tests:

python -m pytest -q

Hosted-safe smoke tests:

$env:WEBOTS_KIT_RUN_SMOKE='1'
python -m pytest -q -k mcp_tool_list_smoke

Full runtime smoke tests with real Webots execution:

$env:WEBOTS_KIT_RUN_RUNTIME_SMOKE='1'
python -m pytest -q -k "session_start_inspect_stop_smoke or benchmark_smoke"

The self-hosted GitHub workflow for runtime smoke expects a Windows runner labeled interactive-webots. That runner must be started from an interactive user session. Windows service mode is not sufficient for Webots runtime smoke because Webots exits during OpenGL initialization before controllers connect. The self-hosted runtime workflow also auto-triggers for runtime-affecting changes and still supports manual dispatch. Packaging and release verification are handled by GitHub workflows:

• Packaging CI
• Release

Troubleshooting

• If doctor fails, ensure WEBOTS_HOME is set or Webots is installed in C:\Program Files\Webots.
• If MCP or session startup closes immediately, inspect session logs for the session artifacts.
• Session failures now report structured error codes such as render-init-failed, agent-connect-timeout, and supervisor-connect-timeout.
• If package installation changes global Python web dependencies, recreate a dedicated virtual environment and reinstall there.
• GitHub-hosted windows-latest runners are only used for unit tests, doctor, and MCP handshake smoke. Real Webots runtime smoke is exposed as a separate manual workflow for self-hosted Windows runners with Webots installed.
• Runtime-affecting changes now also path-trigger the self-hosted runtime workflow when the interactive-webots runner is available.
• examples/ contains runnable demo assets; benchmark thresholds and pass/fail logic live in the benchmark registry inside the toolkit code.
• Wheel installs use bundled package-local scenario assets; source checkouts continue to use repo-local examples/.