MCP server

cicpy ships an MCP server so an agent doing schematic-driven layout can look before it draws: render a cell or a finished GDS as an inline image, inspect placement and connectivity, measure what blocks a lane, and run the design repository’s own checks — without a rebuild per question.

Install and run

pip install cicpy[mcp]     # the server framework + render deps
cicpy-mcp                  # run directly, or point a client at it

An MCP client config entry:

{ "cicpy": { "command": "cicpy-mcp" } }

cicpy is technology independent and so is the server: the technology file is always an argument, and DRC runs through the repository’s own make drc, so the rules and tools stay where the repository put them.

The tools

Ask-before-you-draw — each of these answers without a regeneration:

tool	question it answers
layout_guide	the field guide itself — read it first
render	draw a cell from a .cic, returned inline, y up
render_gds	draw a finished GDS the way a layout engineer reads it; top_only leaves placed blocks as outlines
cell_info	placement in a .cic: instances, groups, pitches, ports
netlist_info	which devices connect to which nets — read before choosing placement groups
stackorder	which columns are interleaved on a terminal, and what reordering buys
tracks	which routing tracks are occupied, and by what; free reports usable spans
blockers	what stops a net dropping a via column in a box — the via-COLUMN check no same-layer scan can do
findroute	search a path for a net and report it without drawing
route_options	what each routing option means, and which ones lie to you
checkroutes	shorts and opens in a built .cic, with BRIDGE/CHAIN attribution
connectivity	rerun the repo’s sch2mag with the connectivity check — the same analysis the GUI shows
drc	run the repository’s own make drc and report the rules that fired

The routing loop these serve is the one in the field guide: measure (tracks, blockers, netlist_info) → search (findroute, or a declared maze) → import the emitted story into the sidecar → verify (drc, checkroutes, plus the repo’s kdrc/lvs/ant targets) → next net.

Two habits that pay:

• blockers before believing a clean track report. A trunk on M4 and a pin on M1 never share a track; what collides is the via column. Every routing failure measured so far was this.
• checkroutes is blind to raw contact paint (paint-stitched supply straps): for those, magic extraction and LVS are the truth.