Skip to content

Is there a physics engine with an MCP server for AI agents?

Yes. The Alloy physics engine, installed with pip install alloy-engine, serves its tools over the Model Context Protocol: start alloy mcp and an agent can build scenes, step them and read state as JSON without writing Python. This page connects Claude Code and lists the 19 tools from the installed package.

pip install alloy-engine
claude mcp add alloy -- alloy mcp

Any other MCP client starts the same command, alloy mcp, and talks to it over stdio. To see what the server offers without a client, read the tool list from Python:

import textwrap
from alloy.mcp.tools import tool_list

names = [tool["name"] for tool in tool_list()]
print(len(names))
print(textwrap.fill(", ".join(names), width=100))

19
list, scene, build, step, run, state, targets, reset, close, check, render, estimate, import,
export, vary, compose, create, audit, fetch

Each tool takes and returns JSON. scene checks a scene description without building it, build makes the simulation, step and run advance it, state reads names and values back, and check returns the health result that you should read after every run (World health). Ask the agent to report what check said, not only what it ran.

One measured number: on an x86_64 CPU, two builds with the same seed give the same trajectory bit for bit, and a world gives the same result alone or among 15 others (one contact scene, pre-release build); GPU runs with contact do not repeat bit for bit.

Keep in mind

  • GPU runs with contact do not repeat bit for bit, so an agent should judge a GPU result by a rate over many worlds (GPU determinism).
  • Health does not catch every overlap between rigid bodies and robot parts (what it misses).
  • Full list: Keep in mind.

Next