SimTooReal

Build, verify, and deploy robot policies with Claude Code, Codex, Cursor, or any AI agent for creation of robots as you desired with our mcp — leave the rest from first reward function to safe hardware rollout to us.

SimTooReal is the complete robotics development platform for the agentic AI era. It connects simulation training to physical hardware deployment through an evidence-backed pipeline, with a runtime physics safety governor sitting at the boundary between policy output and robot execution. Every stage — authoring, training, diagnosis, verification, governing, deployment — is accessible to both human engineers and AI coding agents through a consistent toolchain.

This is not a generic MLOps dashboard adapted for robots. SimTooReal is built ground-up for robotics, with first-class support for Isaac Lab (NVIDIA Isaac Sim / PhysX), MuJoCo, MuJoCo MJX (GPU-batched via JAX), Newton/MJWarp, and LeRobot VLA workflows. Internal tooling handles the simulator integrations so you don't have to wire them yourself.

An AI Agentic Stack for Robot Creation SimTooReal ships 8 MCP (Model Context Protocol) lifecycle tools that wire directly into Claude Code, GitHub Copilot, Codex, Cursor, Windsurf, and any MCP-compatible agent. Your AI coding agent can list and init robot projects, push a local project folder, dispatch training to SSH hosts or managed GPUs, inspect live reward curves and failure clusters, fetch deployment readiness reports, and read governor verdicts — all as structured context, not log noise.

The reward iteration loop that used to require an engineer reading dashboards can now be agent-driven: write a reward function, push the project, the agent reads the metric stream, suggests a change, relaunches. SimTooReal keeps full run lineage — every hyperparameter change, checkpoint, and readiness verdict — so the agent has evidence to act on.

CLI — install from npm, stays deliberately thin:

npm install -g simtooreal-cli npmjs.com/package/simtooreal-cli

Project push, remote training dispatch, log streaming, status, and run history all route through the same backend as the web app and VS Code extension.

VS Code Extension — full project and training lifecycle from inside your editor, powered by the same MCP tools available to Claude Code and other agents:

SimTooReal for VS Code: https://marketplace.visualstudio.com/items?itemName=sim-too-real.simtooreal-vscode

The Platform Author — Start robot projects from the CLI, Studio, VS Code, or an MCP agent call. Templates cover locomotion robots (Unitree Go2, G1, H1, Anymal-C/D, Spot) and manipulation (Franka Panda, custom URDF/MJCF/USD). Curriculum configs, domain randomization presets, hyperparameter sweeps, checkpoints, and training videos stay attached to every project permanently.

Diagnose — SimTooReal groups crashes, drift episodes, action saturation, and gait failures into a single failure-intelligence view. Automatic clustering, rerun recordings, metric timelines, and root-cause suggestions let you fix a failure mode while the run is still correctable — not after hardware testing has already failed.

Verify — Policies are scored across three independent physics engines — PhysX, MuJoCo/MJX, and Newton/MJWarp — before any promotion. A five-pillar readiness score covers observation match, action coverage, physics alignment, robustness, and recovery readiness. Thirteen deployment checks combine cross-engine confidence, safety envelope results, checkpoint integrity, and run evidence into a single pass/block verdict.

Govern — TheWorld is the runtime physics safety governor between policy output and hardware execution. It classifies contact risk using a KV + Hertz contact model across six certified material classes, computes force limits, and runs a fast 120Hz veto loop at 2.4µs latency. Block verdicts feed directly into deployment gates — an unsafe policy cannot be promoted regardless of reward score. Governor traces are exposed as MCP context so your AI agent can read the veto reason and adjust accordingly. Architected to align with EU Machinery Regulation 2027 for collaborative robots.

Deploy — Saved models are tied to readiness certificates, governor verdicts, and unit telemetry baselines. Canary rollouts, pause controls, and rollback are first-class. Fleet telemetry feeds drift gates so field degradation triggers a retrain loop automatically.

Pricing Tier Includes Free Live metrics, failure intelligence, CLI, MCP tools, MuJoCo + LeRobot, Isaac Lab, cross-engine validation, readiness gates, physics cockpit, unlimited runs Enterprise Fleet deployment, staged rollout, custom GPU fleet, SSO, SLA Simulators: Isaac Lab · MuJoCo · MuJoCo MJX · Newton/MJWarp · LeRobot Safety: TheWorld governor Try it: https://www.simtooreal.com simtooreal-cli : https://www.npmjs.com/package/simtooreal-cli vs code extension : https://marketplace.visualstudio.com/items?itemName=sim-too-real.simtooreal-vsco