sim-cli lets AI agents operate CAE solvers one verified step at a time.

sim is an open-source CLI and local runtime that lets Codex, Claude Code, GitHub Copilot, Gemini, and other agents work with simulation software through solver-specific plugins and bundled skills. An agent can check what is installed, connect to a solver, inspect live state, execute bounded modeling or analysis steps, verify result/state, and save checkpoints/artifacts for engineering review.

Quick Start · COMSOL + Codex · Agent Loop · Remote Solvers · Plugins · Commands

---

Who this is for

sim is for agents and people trying to get real simulation work done.

• CAE engineers who already script solvers and want an agent to help automate COMSOL, Abaqus, HFSS, Fluent, MATLAB, LTspice, and similar tools without losing inspection and recovery between steps.
• Design engineers and occasional simulation users who have agent experience and want an agent-assisted workflow: ask for a simulation, watch the model evolve, review screenshots or plots, and keep final artifacts.
• AI agents reading this repository to learn the safe setup and operating loop before touching a solver.
• Engineering leaders evaluating whether agent-assisted simulation can be repeatable, reviewable, and compatible with existing solver installations.

Plugin authoring, runtime internals, and driver protocol details live in docs/DEVELOPMENT.md.

What sim gives an agent

LLMs can often write solver scripts, but a one-shot script is a weak workflow: it hides intermediate state, fails late, and makes recovery difficult.

sim gives an agent a small, composable control surface: detect the environment, attach to a live session, inspect state, run one bounded step, verify, and checkpoint. Use solver-native batch commands directly when they are the right execution primitive.

A bounded CAE step is one small modeling, meshing, solving, or postprocessing action that can be inspected and verified before continuing. Examples: create a geometry feature, assign a material, apply a boundary condition, generate a mesh, run one study, extract a probe value, create a plot, or export a result table.

The solver-specific knowledge is not baked into the core CLI. It comes from plugins. A plugin can provide both:

• a driver, so sim can launch or talk to the solver
• a skill, so the agent knows the solver-specific workflow, pitfalls, and inspection rules

Human-in-the-loop collaboration

sim is designed for shared control, not unattended black-box automation. When a solver plugin exposes live state through sim inspect, the agent can re-read the current solver session after each meaningful step. That means an engineer can cut in through the solver GUI, change geometry, parameters, boundary conditions, plots, or saved artifacts, then ask the agent to inspect again and continue from the real current state.

This is the collaboration model: the human can watch, correct, and steer; the agent keeps using inspection and checkpoints instead of assuming its previous script still matches the real solver state.

Quick Start: agent setup

Use this path when the agent and solver are on the same machine. You do not need to start sim serve manually for the local happy path; sim connect will use the local runtime. The default docs use uv so agents run the sim and plugins declared by the current project instead of guessing which executable is on PATH.

uv run sim ... runs sim from this project environment, so it sees this project's installed solver plugins. Run from the project root:

uv init  # only if this is not already a uv project
uv add sim-cli-core sim-plugin-comsol
uv run sim plugin sync-skills --target .agents/skills --copy
uv run sim check comsol
uv run sim plugin doctor comsol --deep

The sync-skills target depends on your agent: use .agents/skills for Codex and GitHub Copilot, or .claude/skills for Claude Code. Substitute that path in any sync-skills command shown below.

Without uv

If you cannot use uv, create a normal Python virtual environment, install sim-cli-core and the solver plugin into that environment, then run sim from the activated environment:

python -m venv .venv
source .venv/bin/activate   # Windows PowerShell: .\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install sim-cli-core sim-plugin-comsol
sim plugin sync-skills --target .agents/skills --copy
sim check comsol
sim plugin doctor comsol --deep

Hand the task to your agent

After setup, give your agent the engineering goal in plain language — the simulation you want and the quantity you want reported. You do not need to recite the operating loop: the bundled solver skill already enforces the inspect / verify / checkpoint discipline once the agent loads it.

Two things are still worth telling the agent explicitly:

• Run sim through this project with uv run sim ... so it sees the project's installed plugins.
• Don't guess solver API names — inspect the live model or the solver's local docs first. If you change something manually in the solver GUI, ask the agent to re-inspect live state before continuing.

A concrete example follows below.

Example: COMSOL and Codex on one machine

Follow Quick Start to install sim-cli-core and sim-plugin-comsol and sync the skill — those commands already use COMSOL. Then ask Codex for the actual task:

Simulate the natural-convection cooling of attached `pcb.mph` and report the
maximum junction temperature. Use the installed COMSOL skill. Check COMSOL with
`uv run sim check comsol` first. If you need a visible live COMSOL Desktop
session, connect with `--ui-mode gui`.

The prompt names a goal and a quantity to report — not a list of sim sub-steps. The COMSOL skill supplies the rest: session health and model identity checks, checkpoint policy, and step-by-step build/solve discipline.

For COMSOL-specific details such as shared Desktop mode, offline .mph inspection, Desktop attach fallback, model identity checks, and checkpoint policy, follow the bundled COMSOL skill.

Why CLI-first?

Engineering simulation is file-based, scriptable, local, artifact-heavy, and long-running. CAE agents work with solver executables, model files, Python/Java/journal scripts, shell commands, logs, checkpoints, and plots — a CLI command surface composes with all of that and matches how Codex CLI, Claude Code, and other coding agents already operate.

MCP is useful for API-style integrations and remote tool discovery, but a broad MCP surface adds context overhead and wrapper maintenance. For COMSOL, Abaqus, Ansys Workbench, OpenFOAM, LTspice, and similar solvers, sim-cli keeps the source of truth as a small, auditable command loop.

The agent loop

For any solver, the agent should prefer this loop over one large generated script:

1. uv run sim check <solver> to detect installed solver versions and plugin compatibility.
2. uv run sim connect --solver <solver> ... for live work, uv run sim run when its wrapper adds value, or a solver-native batch command when better.
3. uv run sim inspect session.versions and the solver-specific health or identity target before changing state.
4. uv run sim exec --file step.py --label <step> for one bounded modeling or analysis step.
5. uv run sim inspect last.result and solver-specific state.
6. Verify the result/state using solver-specific numerical evidence when available.
7. Save checkpoints and artifacts when the solver plugin or skill requires them.
8. uv run sim disconnect when the session is done.

Screenshots and plots help humans review the result, but engineering acceptance should prefer numeric evidence when the solver skill defines it: mesh statistics, convergence, finite probes, conservation checks, tolerances, or expected trends.

Local vs remote solvers

Same machine: install sim-cli-core and the solver plugin into the project environment, sync the skill to your agent, then use uv run sim connect. Do not add --host unless you are intentionally talking to a remote sim serve.

Remote solver workstation, lab box, or HPC login node: install sim-cli-core and the solver plugin on the solver host, start sim serve there, then point the local agent at that host:

# On the solver host.
uv run sim serve --host 0.0.0.0 --port 7600

# On the agent/control machine.
uv run sim --host <solver-host-ip> check <solver>
uv run sim --host <solver-host-ip> connect --solver <solver>
uv run sim --host <solver-host-ip> inspect session.summary
uv run sim --host <solver-host-ip> disconnect

Only bind sim serve to a trusted network such as a VPN, Tailscale, or a protected LAN. The runtime currently has no auth layer, and /connect plus /exec can execute solver-side code.

Solver plugins

sim-cli-core ships with no solver drivers built in. Each simulation solver is reached through an explicit plugin package.

A few representative plugins:

Solver	Package spec	Plugin repo
COMSOL	sim-plugin-comsol	sim-plugin-comsol
Abaqus	sim-plugin-abaqus	sim-plugin-abaqus
Ansys Workbench	sim-plugin-workbench	sim-plugin-workbench
Autodesk Fusion 360	in development	sim-plugin-fusion360
LTspice	sim-plugin-ltspice	sim-plugin-ltspice

For the curated full list, see sim-plugin-index.

After adding any plugin package, sync its bundled skill and verify that the local solver can be reached:

uv run sim plugin list
uv run sim plugin sync-skills --target .agents/skills --copy  # or .claude/skills for Claude Code
uv run sim check <solver>
uv run sim plugin doctor <solver> --deep

For direct wheel, Git, local checkout, or non-uv package workflows, see docs/plugin-install.md.

Project setup with sim.toml

For reproducible Python packages, commit the pyproject.toml and uv.lock created by uv add. Use sim.toml for solver defaults and workspace settings:

uv run sim init

Example:

[sim]
default_solver = "comsol"
workspace = "./workspace"

[[sim.plugins]]
name = "comsol"
package = "sim-plugin-comsol"

Then a fresh checkout can run:

uv sync
uv run sim setup --dry-run
uv run sim plugin sync-skills --target .agents/skills --copy  # or .claude/skills for Claude Code

Common commands

Command	Use it for
uv run sim plugin list	Show plugins visible in this project environment.
uv run sim plugin info <solver>	Show plugin metadata and compatibility summary.
uv run sim plugin doctor <solver> --deep	Check plugin wiring plus local solver detection.
uv run sim plugin sync-skills --target .agents/skills --copy	Materialize installed plugin skills for your agent (.claude/skills for Claude Code).
uv run sim check <solver>	Detect local or remote solver installs.
uv run sim connect --solver <solver>	Open a persistent solver session.
uv run sim exec --file step.py	Run one bounded step in the live session.
uv run sim inspect <target>	Query session, result, or solver-specific state.
uv run sim run script.py --solver <solver>	Run a deterministic one-shot script.
uv run sim disconnect	Tear down the active session.
uv run sim setup	Validate sim.toml and report declared plugin package specs.

Run uv run sim describe for a machine-readable command manifest, or uv run sim <command> --help for exact options.

Solver ownership

sim-cli does not bundle or redistribute simulation solvers or vendor SDKs. Install and operate each underlying solver according to its vendor terms. See NOTICE for optional SDK dependency notes.

sim-cli is an independent open-source project and is not affiliated with, endorsed by, or sponsored by any solver vendor. Product, solver, and company names remain the property of their respective owners.

Developer docs

• docs/DEVELOPMENT.md - project setup, layout, driver development, and architecture notes
• docs/plugin-install.md - plugin installation reference

License

Apache-2.0 - see LICENSE.