wobbl 0.1.0

Simulate anything before you build it.
Verified strength, heat and circuit simulation for coding agents and the people who design physical things.

Website  |  Getting started  |  Guides  |  CLI reference  |  Examples

_{A real Wobbl output, a sand cast motor mount under a 2 kN belt pull, with the peak of 124 MPa at the rib tip and a safety factor of 1.33 against a target of 2.}

What it is

Wobbl is a command line tool, an MCP server and an agent skill that run real engineering solvers and check every answer before anyone acts on it. Point it at a STEP, IGES, STL or parametric Python part, ask a question in plain words or with explicit loads and supports, and you get back the number, where it happens, a confidence level with its reasons, the assumptions, a heatmap with the failure point marked and fixes with numbers.

• Strength with CalculiX, on second order tetrahedral meshes from gmsh and OpenCASCADE that are refined around the peaks until the stress settles.
• Heat, steady and transient, also with CalculiX.
• Circuits with ngspice, for LEDs, RC networks, buck converters, MOSFETs or any netlist.

wobbl fix --apply edits the part, adding fillets and gussets to any STEP file, and runs the whole study again. The same tools work from your terminal and from Claude Code, Codex, Cursor and Gemini CLI.

Quick start

Wobbl needs Python 3.10 or newer, and until it is on PyPI you install it from GitHub.

pipx install "git+https://github.com/sohan-bhat/wobbl"
# or: uv tool install "git+https://github.com/sohan-bhat/wobbl"

wobbl setup     # solvers into ~/.wobbl, MCP server and skill for your agents
wobbl doctor    # check every piece

wobbl setup downloads a pinned micromamba and installs the solvers from conda-forge into ~/.wobbl (about 340 MB, no admin rights). Use wobbl setup --no-agents for solvers only, or --agents claude,cursor to pick.

This first answer needs no files.

wobbl "will a blue LED on 5V with a 47 ohm resistor fry?"

Then try your own part.

# holes, fillets, sharp corners and face ids
wobbl inspect part.step

# ask in plain words, or give the material, supports and loads yourself
wobbl part.step "will this 6061 bracket yield under 50 lbf at the tip?"
wobbl run part.step -m petg --fix "bolt holes d=5.5" \
  --load "250 N down at 0,-15,31 @ holes d=3.4 rigid"

# build the fixes (fillets, gussets, thicker walls) and run each again
wobbl fix wobbl-out/part --apply

Or run a bundled example, which takes a few minutes.

git clone https://github.com/sohan-bhat/wobbl
cd wobbl/examples/shelf-bracket && ./run.sh

Example output

This is a printed PLA shelf bracket, screwed to a wall and carrying its share of an 8 kg shelf. Every run writes summary.txt, which is what an agent reads, along with report.html, outcome.json, heatmaps and result.vtu for ParaView.

$ cd examples/shelf-bracket
$ wobbl run bracket.py -m pla --fix "bolt holes d=4.5" --fix "wall x=min" \
    --load "40 N down @ top and within 12 of 65,12.5,6" -o output
$ cat output/summary.txt
FAILS. Peak 52.3 MPa on the 0.5 mm inside fillet (face 3) at (6.19, 12.6, 6.11). The part cracks at about 0.96x this load.

Confidence: high. passed mesh convergence, force balance, hand calculation, singularities, bolt bearing, units and sanity
Max deflection 3.47 mm at (90, 0, 6).

Checks:
  [pass] mesh convergence: max principal stress settled: 43 -> 50.9 -> 53 -> 52.3 MPa over 4 meshes, last change 1.4%
  [pass] force balance: reactions balance the loads to 0.00014% (forces) and 8.4e-06% (moments)
  [pass] hand calculation: beam theory through the critical section gives 15.2 MPa nominal; FEA peak is 3.45x that, a stress concentration of about 3.4 from the local shape
  [pass] singularities: no sharp corner or support artifact is above the reported stress
  [pass] bolt bearing: worst hole of fix1 (d 4.5 mm) carries 2.7 N sideways and 45.7 N along its axis; bearing stress 0.1 MPa, safety factor 499.33 on bearing
  [pass] units and sanity: part 123 mm across, peak strain 0.015, max deflection 3.47 mm (2.8% of size): all plausible

Suggested fixes:
  - Stress piles up in the 0.5 mm fillet. A fillet of 3 mm (about half the 6 mm wall) spreads it out. ...
  - The peak sits where two walls meet and bend. A gusset as thick as the wall ...

wobbl fix output --apply then builds each fix and solves it again.

An agent on its own versus Wobbl

Given a CAD file, a strength question and 15 minutes, a capable coding agent working alone answered in 1 of 10 trials across six parts. Its hand calculations on beam like parts were good, within 1 to 5%. On the ribbed cast motor mount and the thin walled enclosure, the two parts that actually fail, it delivered nothing in six trials, because it kept writing its own finite element code and hitting the same traps, such as meshers folding thin curved walls, singular peaks at bonded bolts and load idealizations that flip the verdict.

Wobbl answers the same parts in minutes and handles each of those traps by default.

The full comparison, with times and every trap, is on the website.

Using it from agents

wobbl setup registers an MCP server and installs a skill for every coding agent it finds, backing up each file it touches. wobbl uninstall removes exactly what it added.

The agent gets 13 tools, inspect_geometry, mesh, new_study, set_material, apply_load, add_fixture, solve_structural, verify, render, suggest_fix, solve_thermal, solve_circuit and list_materials. Results are compact JSON with image and report paths. The skill teaches the workflow and what never to skip, which means reading the checks before the number, never quoting an excluded peak and making the supports match the real build.

> Will this printed PLA shelf bracket hold its share of an 8 kg shelf?

agent: solve_structural(path="bracket.py", material="pla",
                        fixtures=["bolt holes d=4.5", "wall x=min"],
                        loads=["40 N down @ top and within 12 of 65,12.5,6"])
       verdict fails, 52.3 MPa on the 0.5 mm inside fillet, safety factor 0.96, confidence high
agent: suggest_fix(result_dir, apply=true)
       3 mm fillet 24.1 MPa (2.08), legs at 10 mm 21 MPa (2.38), gusset 9.82 MPa (5.09)

Strength studies take one to seven minutes on a 2018 laptop. Each agent's setup, the timeouts and every tool are on the website.

What it checks

Every strength answer runs these checks and turns them into a confidence level, high, medium or low, always with the reasons. A result that failed a check is never reported as high confidence.

• Mesh convergence over at least three meshes, refined around the peaks that matter.
• Force and moment balance, with the reactions cancelling the loads to 0.5%.
• A hand calculation with beam theory through the critical section, and a check that the stress concentration is physical.
• Singularities, with sharp inside corners found from exact CAD normals and the peaks there and at idealized supports excluded, named and tracked.
• Load introduction, where off part loads are solved both rigid and distributed, each converged at its own peak, and the spread is reported.
• Bolt bearing at bolted holes, in place of the idealized peak.
• Units and sanity for the part size, peak strain and deflection.
• Large deflection, a nonlinear re-solve when the part bends a good fraction of its wall.

Heat answers check energy balance and a lumped hand calculation, and circuit answers are compared with the textbook and run again at half the time step. Every study has a time budget, and when it runs out the last finished mesh level comes back with its checks. The same input gives the same answer on every run, because meshing, wall thickness sampling and the budget are all deterministic. The website explains every check and how the confidence is set.

Validation

Eighteen cases with known answers run through the same pipeline you get. Every case passes its tolerance, and every case whose reference is exact for the modeled setup lands within 2%.

The full table is on the website. Run it yourself with wobbl validate, which takes about 30 minutes.

Examples

Each folder in examples/ is a real product with its real outputs, the reports, heatmaps and summaries, and each has a page on the website.

Each folder has a run.sh that regenerates its outputs.

Documentation

• Getting started, to install, set up and run a first study
• Guides for strength, heat and circuits, the inputs, the results and fixing parts
• CLI reference with every command and option
• Agents and MCP with what wobbl setup registers, the 13 tools and the skill
• How it checks itself, every check and what sets the confidence
• Validation, the 18 cases and their setups
• Bake-off, a coding agent on its own versus Wobbl
• Contributing, with the architecture and how a new solver plugs in
• Publishing, the release steps for PyPI
• Changelog

Limits

• Materials are linear elastic, so Wobbl reports whether yield is reached but not what happens after, and large deflection is checked with a nonlinear re-solve.
• Loads are static, and fatigue, impact and vibration are not modeled yet.
• Contact is limited to rigid walls the part can push on but not pull, and bolted joints are modeled as clamped holes without preload.
• Each study uses one material, and assemblies of touching bodies are bonded.
• Thin walls are slow, since 2 to 3 mm shells on 100 mm parts need large tetrahedral meshes and take several minutes, and shell elements are on the roadmap.
• Printed plastics are treated as isotropic at 100 percent infill, with a layer strength knockdown when you give the print direction.

Shell elements, fatigue estimates and bolt preload come next in 0.2, then CFD with OpenFOAM in 0.3 and motion with MuJoCo in 0.4.

Contributing

Issues and pull requests are welcome, and this is how to work on Wobbl itself.

git clone https://github.com/sohan-bhat/wobbl && cd wobbl
python -m pip install -e ".[dev]"
python -m pytest -m "not solver"      # fast unit tests
wobbl setup --no-agents               # CalculiX and ngspice
python -m pytest -m solver            # end to end tests with the solvers

A new solver plugs in as one adapter class, as contributing explains. If you change the physics, add or update a case in src/wobbl/validation/cases.py so wobbl validate covers it.

License

MIT, copyright Sohan Bhat.

sohan-bhat.github.io/wobbl