Specialized MCP servers that give LLMs the ability to run quantum simulations, analyze material properties, and access structured knowledge — reducing token usage and dramatically increasing accuracy.
LLMs are powerful but they hallucinate physical constants, guess simulation parameters, and can't run actual computations. MCP servers solve this.
Instead of guessing material properties, the model calls a tool that runs the actual simulation or queries validated data. No hallucination.
A single tool call replaces thousands of tokens of context. The model asks "what's the band gap of X?" — the MCP server returns the exact number.
Run Quantum ESPRESSO, LAMMPS, and DFT workflows directly from the chat. From atomic structure to material properties in one conversation.
MCP is provider-agnostic. Use Claude, GPT, Kimi, local models — the tools work the same way through JSI's orchestration engine.
JSI's master LLM knows which MCP tools each worker has access to. It can route a simulation task to the right worker with the right tools automatically.
Complete workflows: define a crystal structure, run simulations at multiple levels of theory, extract properties, compare with literature — all in one session.
Purpose-built tool servers for computational materials science, designed to plug into JSI and any MCP-compatible platform.
A computational materials science MCP server that gives AI agents the ability to set up, run, and analyze atomistic simulations. From crystal structures to material properties — MagicDory handles the entire workflow so the LLM doesn't have to guess.
A graph knowledge base for materials science — a structured, queryable repository of material properties, crystal structures, simulation parameters, and literature references. Acts as an extension of JSI's Alexandria memory system, but specialized for the physical sciences.
"Calculate the band gap of GaN in the wurtzite structure using DFT."
Plans the workflow: build structure → query knowledge base for parameters → run QE → parse results. Delegates to workers with MagicDory tools.
Returns validated pseudopotentials, recommended cutoff (80 Ry), k-mesh (8×8×4), and lattice parameters from literature.
Generates input files → runs SCF relaxation → runs band structure calculation → parses output. Returns: band gap = 1.71 eV (GGA-PBE), with convergence data.
"The calculated band gap of wurtzite GaN is 1.71 eV (PBE), which underestimates the experimental value of 3.4 eV — typical for GGA. Want me to run a hybrid functional (HSE06) for a more accurate result?"
We build AI-powered tools for computational materials science and engineering. JSI and its MCP ecosystem are part of our mission to make atomistic simulation accessible to every researcher and engineer — not just HPC specialists.
Visit apeironai.eu