AI agents

Any AI writes the model. Vision proves it.

vision mcp is a server of the Model Context Protocol, the standard through which AI agents call tools: Claude Code, Claude Desktop, Cursor, VS Code (Copilot agent mode), Zed and others. With it, an agent writes .vis, checks it with the compiler of Vision, reads what the program does, and runs it, before anyone relies on the result.

Tool	What the agent gets
vision_language_guide	the language: structure, rules of kernels, types, every built-in function, patterns and complete examples
vision_check	ok, or the error with its line, its code (E0423), what it means and a corrected example
vision_explain	the program as formulas, the warnings of physics and modelling mistakes, and, given the request, the numbers and names it does not use
vision_explain_error	the entry of any error code
vision_run	the measures of a run (with their series and distributions over scenarios) and what the program prints

vision_run reads and writes files only inside the project folder (--root, by default the folder where the server starts): absolute paths and .. are refused. On the numpy reference a run has a limit of work; start the server with --device nvidia (or another GPU) for larger models.

Setup

Claude Code:

claude mcp add vision -- vision mcp --root .

Cursor (.cursor/mcp.json), VS Code (.vscode/mcp.json, with "servers" instead of "mcpServers"), Claude Desktop (claude_desktop_config.json):

{ "mcpServers": { "vision": { "command": "vision", "args": ["mcp", "--root", "/path/to/project"] } } }

Use the full path of vision if it is not on the PATH (for example /path/to/project/.venv/bin/vision).

What it gives, measured

The copilot evaluation has 24 basic tasks and 15 hard ones. The hard ones cover vague requests, programs with data files and 2-D/3-D coupling, analytical results, traps and edits of given programs. Every task is checked against physics, not against a reference program.

	Result
Claude Haiku 5.5 (the smallest current Claude model), basic tasks	23 of 24 (the 24th right, under another name)
The same, hard tasks, with the repair loop	14 of 15 in each of the last two runs (a different miss each time)
The same as an agent with the tools of vision mcp, the three trickiest hard tasks	3 of 3
Local models of 0.5 to 1.5 billion parameters	0 of 24

The mistakes left are the dangerous kind: programs that compile, run and look right. Examples are a diffusion between two materials that does not conserve the flux, or a transfer that creates water. The warnings of vision explain and the checks of the run catch them, and the agent corrects itself.

Tested with the official MCP Inspector client, as well as the test suite.