Connect an external agent through MCP

Build/install the native project following the installation guide, then install the required extras from the same checkout:

python -m pip install '.[mcp,cad,results,visualization]'
sphinxsim capabilities --json

Configure an MCP client with the installed native CLI, for example:

{
  "mcpServers": {
    "sphinxsim": {
      "command": "/path/to/venv/bin/sphinxsim",
      "args": ["mcp", "--root", "/workspace"]
    }
  }
}

The server uses the official SDK and stdio transport. It requires no product account, cloud gateway or proprietary handshake. The process owner selects the workspace and supplies resource isolation. Do not expose unrestricted workspaces.

Discover the running inventory instead of copying a stale schema. capabilities reports source evidence and limitations; compile returns a new configuration; validate checks an explicit config_file; project_inspect returns the resource inventory; submit requires that reviewed project inventory. Use submission_id for retries. geometry defaults to a persistent CAD job. All job operations take the returned complete reference object, including its request hash.

A native desktop is optional. Start it with sphinxsim desktop serve ... and give its owner-only reference file to sphinxsim mcp --root ... --desktop-reference .... Only then is desktop advertised. Observe status, then send the explicit instance/version and operation scope; the server never silently fetches a fresh version to make a stale edit succeed. Remote computer-use drivers are external.

Verified artifacts use sphinxsim://jobs/{job_id}/artifacts/{name} resources. Reads are limited to published files and 1 MiB per resource; use owner-provided file delivery for large assets. Resource access rechecks published size/hash evidence.

Read the native workflow guide for CLI examples, JSON contracts, execution lifecycle, quality interpretation and test boundaries. The optional skill guides experimental decisions; install it through your client's supported skill mechanism. It grants no access.

No Docker image is required by this protocol. Image names, GPU devices, remote desktop servers, browser embedding and authentication are deployment choices outside this repository's agent implementation.