Structures, heat, flow, fracture, plasticity — finite-element analysis as one API request. Your agent describes the problem; the engine solves it and returns the answers.
physicsbase/solve over REST, MCP, or Python.# the agent sends a structure { "nodes": [[0,0],[1,0],[2,0],[3,0]], "materials": {"steel":{"E":210e9,"nu":.3}}, "elements": [ {"type":"beam2d","nodes":[0,1], "section":{"A":.01,"I":8e-6}} … ], "supports": [{"node":0,"dofs":["ux","uy","rz"]}], "loads": [{"node":3,"fy":-5000}] } # and gets the solved answer back ← tip_deflection: -26.8 mm ← wall_moment: 15.0 kN·m ← solve_ms: 1.2
No key, no signup, no glue code. The service publishes the standard discovery files agents already look for, exposes an OpenAPI schema, and runs a remote MCP endpoint — so ChatGPT, Claude, and agent frameworks can find it and start solving on their own.
/llms.txtthe agent's map of the API/capabilitieselements, analyses, DOFs — read first/openapi.jsonfull schema for OpenAPI tools/mcpnative tool use over streamable HTTP/solvesend a problem, get answers back# 1. discover — read the map, then the capabilities GET /llms.txt GET /capabilities → 14 elements, 8 analyses, no auth # 2. solve — one POST, no key POST /solve { "nodes":[[0,0],[3,0]], "elements":[…], "supports":[…], "loads":[…] } ← { "displacements":[…], "reactions":[…] } # …or connect it as a remote MCP server "physicsbase": { "url": "https://www.physicsbase.dev/mcp" } → tools: fem_solve, fem_modal, fem_buckling, fem_field …
Handing an agent FEA code to run elsewhere means it also has to derive the formulas — and it will hallucinate them. physicsbase moves the physics server-side. The agent only describes geometry and loads; every number it reasons about is computed, not guessed.
PL³/3EI, hopes it's right"loads": [ {"node":3, "fy":-5000} ]
"reactions":[ {"fy":5000, "mz":15000} ]
A full structural FEA curriculum plus a unified field kernel for heat, diffusion, and flow — and the couplings between them.
Springs, trusses, 2D & 3D frames, higher-order continuum, solids and axisymmetric — plus a unified field engine for heat conduction, mass diffusion, potential flow, and transport. All in one JSON format.
closed-form tests, all green
One JSON format, three ways to call — however your agent is wired.
Frequencies, transient response, damage, and von Mises plasticity.
Critical load factors from geometric stiffness — columns & frames.
Point, distributed, thermal, gravity, self-weight, tractions.
Member forces, reactions, deflections, moment diagrams — 2D and 3D, with torsion and biaxial bending. Tension and compression resolved sign-correct.
Mesh a 2D part and get the full stress field with von Mises per element, plane stress or plane strain, linear or quadratic elements so an agent can find the critical spot.
The shapes trace the real deformed states an agent gets back — a beam bends, a mode vibrates, a truss trades tension for compression.
static · deflection & moment
modal · natural frequency
tension & compression
The same engine that bends a beam solves temperature fields, potential flow, and nonlinear damage — here's what those look like moving.
a source spreads through a conductive field
streamlines bending past an obstacle
stiffness degrades, a crack propagates
We took hard problems straight from arXiv and solved them by calling physicsbase in a loop — the results match the papers and the closed-form theory.
Every frame is a fresh finite-element solve on the same mesh — only the crack grows. This is the eXtended FEM reproducing the classic edge-crack benchmark, with the stress-intensity factor landing within 1.5% of the handbook and converging as the mesh refines.
Read the case study →
Phase-field brittle fracture — coupling the structural and field solvers to propagate a crack through a notched specimen.
Read the case study →
The eXtended FEM — a crack cutting through elements with no remeshing. K₁ within 1.5% of the handbook, and converging.
Read the case study →
SIMP topology optimizat