Quickstart

Sign in and land on /chat. Chat is the primary surface — projects, cases, runs, and reports are durable deep links that hang off it. There is no wizard. There is no case form to fill in. You type what you want to model.

State the physics, geometry source, quantities of interest, and credibility bar. A good first message names what and why, not how. Example:

Evaluate airflow through this electronics enclosure using the attached CAD. Goal: keep the outlet temperature rise under 12 K and report pressure drop, residuals, and uncertainty assumptions.

The agent inspects the workspace, searches local solver knowledge, loads skills if relevant, and may ask one or two clarifying questions before drafting a typed plan.

The agent emits a typed plan: solver choice, mesh intent, boundary conditions, controls, validation criteria, and an estimated cost envelope. This becomes the seed of a CaseSpec. You can ask for changes in chat or accept and continue. Until you approve, no compute runs.

When the agent is ready to write files and execute commands, it surfaces a typed approval envelope: operation, cwd, command, blast radius, and reuse scope. Approve once, and subsequent same-shape commands inside the case may reuse the approval. Sensitive operations (publishing, sharing, deleting) always re-prompt.

The live-monitor sidecar tails solver logs and parses residuals, iteration counts, and monitor fields in real time. The case view shows residual plots, mesh metrics, and any failure cards. Diagnose-and-recover happens in the same chat thread; the agent inspects logs and proposes a fix without losing context.

Every result points at an artifact:

• residuals → solver log, parsed by the solver pack
• QoIs → typed extraction recipe over a named selection
• figures → figure-catalog binding with a renderer and viewpoint
• citations → knowledge-pack chunks with heading paths

If you cannot trace a number to a file, the report cannot publish.

Reports project case state, validation records, figures, standards-pack requirements, and citations into a deliverable. Publishing freezes a content-addressed bundle so another engineer can replay the run, regenerate the figures, or challenge the result.

• Walk the mental model in Workspace, Project, Chat, Case, Run.

• Understand the kernel in The Agent Runtime.

• Browse starter cases — here are a few public templates that compile through typed case state:

• OpenFOAM Backward-Facing Step Benchmark Starter (OpenFOAM) — 4 parameters

• OpenFOAM Lid-Driven Cavity Benchmark Starter (OpenFOAM) — 9 parameters

• OpenFOAM Heated Cavity Starter (OpenFOAM) — 9 parameters

• OpenFOAM Buoyant Room HVAC Starter (OpenFOAM) — 9 parameters

• OpenFOAM rhoCentralFoam Nozzle Starter (OpenFOAM) — 7 parameters

• OpenFOAM interFoam Dam Break Starter (OpenFOAM) — 7 parameters