● A benchmark of real scientific simulations

Scaling simulation
for scientific discovery.

SimulationHub gathers the simulations that scientists actually run in Nature-family papers — across physics, chemistry, biology, materials, climate and beyond — behind the Simulation Context Protocol (SCP), a strict MCP profile, so coding and computer-use agents can select, run, and discover with them.

/ 01 — the hub

Simulations

The full SimulationHub catalog — every entry is a real tool used by scientists or educators.

/ 02b — playground

Simulation playground

Run free simulations with real open models via OpenRouter. A coding agent writes the call for CLI/library sims; a Qwen3-VL + Gemini-3-Flash harness drives the GUI sims by sight.

1 · Models (coding sims use coding models; GUI sims need vision)
2 · Test cases simulation × task
Computer-use agent — live on a GUI simulation raw screenshots · live · the model generates each action and you watch it happen
  Pendulum Lab — live GUI
live action stream
Pick a vision model + GUI sim, then Run live — the model reads the raw screen and acts, step by step, in real time.
/ 02c — the motivation

Are simulations really required?

We read each paper experiment-by-experiment to test whether its science actually depends on a simulation — the evidence behind building SimulationHub.

/ 03 — SCP

Simulation Context Protocol

SCP is a strict profile of MCP (Model Context Protocol): a scientific simulation is an MCP tool, a run is a tools/call, and every scientific extra lives in MCP's own extension points — so any MCP client calls sim #3 exactly like sim #9,000, and SCP-aware clients additionally get selection and the scientific profile.

Simulation = MCP Tool

One Tool descriptor per sim: standard inputSchema/outputSchema, plus taxonomy, cost, conformance, units & provenance carried in MCP _meta under org.simulationhub.scp/.

Result Envelope

Live, mock, programmatic and interactive runs all return the same structuredContent: status, outputs, content-addressed artifacts (MCP Resources), metrics, reproducible provenance, and an is_synthetic taint bit.

Selection & mock

The scp.find meta-tool ranks the whole catalog so agents select tools (RQ2) and recruit out-of-domain ones (RQ3). Paid sims ship a first-class, schema-valid mock so every task runs end-to-end.

Conformance & eval

Guarantee levels L0 manifestL1 liveL2 reproducibleL3 interactive plus a harness-authored verification verdict — so RQ1 use and RQ4 distillation are scorable, not just well-formed.

/ 04 — SCP reference

Simulation Context Protocol

Loading documentation…
/ 04 — provenance

Grounded in real papers

88 Nature-family papers, verified by download, each contributing a benchmark task and the simulations it used.

/ 05 — at a glance

The landscape

Auto-generated statistics over the corpus. Full report in analysis/.