Introduction

SimPilot is an autonomous engineering agent for simulation. You describe an engineering outcome in chat — physics, geometry, quantities of interest, credibility bar — and the agent plans, sets up a typed case, generates solver-native files, runs governed commands, parses evidence, validates against your criteria, and produces a replayable report. Every claim points at an artifact; every action passes through a typed approval envelope.

SimPilot is built for the engineer who would otherwise spend a week wiring OpenFOAM dictionaries, SU2 configs, or CalculiX input decks by hand — and the reviewer who has to trust the result. It is not a wizard, not a black-box prediction service, and not a generic chatbot wrapped around a solver. The agent reads your workspace, navigates local solver resources, follows signed skills, runs the same solvers you would, and shows its work.

Five surfaces, one kernel:

• Chat — primary interface. Plain language in, typed events out (plans, tool calls, approvals, run records, figures, reports).
• Workspace — file tree, terminal, viewer, case state, and validation panels backed by the same typed artifacts the agent sees.
• Cases — typed simulation contracts (CaseSpec, ValidationSpec, StudySpec) that drive solver materialization.
• Runs — governed shell execution producing logs, residuals, artifacts, and cost.
• Reports — evidence-backed deliverables with citations, signatures, and replay manifests.

• Typed state beats prose. The CaseSpec is the source of truth — not the latest chat paragraph.
• No claim without evidence. A QoI value, a convergence statement, a validation verdict — each ties back to a run, parser, artifact, or citation.
• Approvals are typed. File writes, solver launches, sharing, and destructive actions pass through a typed envelope with operation kind, cwd, blast radius, and reuse scope.
• Replay is durable. Every chat turn is a rollout record; every run is reproducible from content-addressed inputs.

Start with the Quickstart to run a real case end-to-end in under five minutes. Skim Workspace, Project, Chat, Case, Run for the mental model. Then move to Core Concepts to understand the kernel, typed state, and approval contracts.