MCP server

The point of cictikz is that an AI assistant can draw a schematic, look at it, and iterate. The MCP server is that loop over stdio.

Registration

pip install -e ".[mcp]"
claude mcp add cictikz -e CICTIKZ_EXAMPLES=$HOME/path/to/figures -- cictikz-mcp

CICTIKZ_EXAMPLES (optional) points at a directory of .tex figures — for example a course repo’s tikz/ — and enables the example tools.

Tools

tool	what it does
`render_tikz(source, body_only=True, dpi=150)`	compile TikZ, return the PNG inline; on failure return the parsed TeX errors
`render_file(path, dpi=150)`	compile an existing figure file (repo-relative `\input` handled)
`list_symbols(query="")`	the symbol library, one line each
`symbol_info(name)`	signature, pin geometry, anchors, example
`style_guide()`	the packaged house rules
`draw_schematic(spec_json, fmt)`	build a circuit structurally from the IR (JSON), emit TikZ or xschem
`tikz_to_xschem(source)`	dialect TikZ body → `.sch` content, connectivity inferred from geometry
`xschem_to_tikz(path)`	any `.sch` → dialect TikZ body
`list_examples(query="")`	search the example corpus by filename and header comment
`get_example(name)`	full source of one example figure

The intended workflow

style_guide() once, so the drawing follows the house rules.
list_examples("cascode") — find an existing figure that looks like the one you are about to draw, get_example it, imitate its idioms.
list_symbols() / symbol_info("lvnmos") for macros and pin names.
Draw, then render_tikz(...) and look at the image — a clean compile says nothing about whether the drawing is right.
For structural work (netlists in, drawings out), build the IR as JSON and use draw_schematic; the docstring carries a complete spec example.

Heavy imports are deferred into the tool bodies, so the server starts fast; rendered images travel as PNG bytes and temporary files are cleaned up per call.