Skip to content

Runtime safety and the guard boundary

aster is the design-side library. Real AEDT solves are executed by the companion guard ansys-aedt-mcp, and the two are designed to be used together with a fail-closed policy.

What the guard enforces (v0.3.2+)

  • Managed sessions only: AEDT sessions launched through the guard are recorded; attaching to a session the guard did not launch is refused. guard_status classifies every live ansysedt.exe as managed or unmanaged.
  • No arbitrary broker code by default: run_python_code / run_python_script are refused at the MCP surface and the worker backend unless a human operator explicitly enables them.
  • Token-scoped solves: when AEDT_GUARD_ENFORCE=1, analysis submissions need a run token bound to project/design/setup; guarded-run adapter commands are audited.
  • Launch/release audit and PyAEDT app identity checks against the verified desktop.

Client rule (mandatory)

If the guard MCP tools (mcp__aedt__* in DeepSeek Harness, mcp__ansys-aedt__* in Codex) are not available, stop. Do not launch ansysedt.exe, do not import PyAEDT, and do not edit .aedt files. Ask the operator to enable the guard. Ready-to-use fail-closed configurations live in the guard repository under examples/clients/ (DeepSeek Harness MCP bundle and Codex config/AGENTS).

aster's trusted bridge

aster executes its modelling snippets through the guard, which refuses run_python_code by default. The aster bridge therefore enables AEDT_MCP_ALLOW_ARBITRARY_CODE only for the duration of its own trusted broker call (the child broker inherits it and the environment is restored afterwards); DSH and Codex MCP servers are separate processes and stay default-deny. Do not co-host aster and DSH/Codex in one Python process: that would share the environment and defeat the scoping. This is an operator-side trusted-tool decision, not something an agent should enable for a client session.

Residual risk

The guard is not an OS sandbox. A same-user process with unrestricted shell can still start AEDT directly; the guard will mark that session unmanaged and refuse to attach, but cannot physically prevent it. For a hard boundary, run the solver under a separate account or VM with AppLocker/WDAC and expose only the guard service. The guard registry and audit log are same-user-writable integrity controls, not a security boundary.

See the guard's THREAT_MODEL and CLIENTS for the full contract.