Layout Canvas

Block-level, AI-native analog layout canvas — parametric block library, typed Block IR, deterministic layout compiler, in-process DRC/LVS/PEX verification, and a local-first web canvas that AI agents can drive through MCP.

Live demo → https://kairos2425.github.io/layout-canvas/

Block IR (JSON) ── compile ──▶ gdsfactory ──▶ GDS
      │                            │
      │                            └──▶ SPICE netlist ──▶ LVS vs extracted netlist
      │                            └──▶ KLayout pya ──▶ DRC subset + extraction
      │                            └──▶ ngspice ──▶ schematic & post-layout sim
      │
      └── constraints · ports · metadata ──▶ MCP tools for AI agents

What it does

Install

pip install layout-canvas  # PyPI — or: pip install git+https://github.com/Kairos2425/layout-canvas
pip install klayout        # in-process DRC/LVS/extraction engine
pytest                     # 245 tests; tool-dependent tests auto-skip

Optional tools (probed adapters — absent tools report unavailable, never fake results):

• ngspice — schematic & post-layout simulation (LAYOUT_CANVAS_NGSPICE=<path>)
• KLayout app / netgen / magic — full foundry decks (LAYOUT_CANVAS_KLAYOUT, …)
• IHP SG13G2 models — LAYOUT_CANVAS_IHP_MODELS, LAYOUT_CANVAS_OSDI_DIR (see docs/EXTERNAL_TOOLS.md)
• Remote Virtuoso/spectre host — LAYOUT_CANVAS_VIRTUOSO_HOST, LAYOUT_CANVAS_VIRTUOSO_DIR, LAYOUT_CANVAS_SPECTRE_MODELS (SSH key auth; see docs/VIRTUOSO_ACCEPTANCE.md)

Quick start

CLI

layout-canvas compile examples/diffamp.json -o diffamp.gds    # IR → GDS
layout-canvas netlist examples/diffamp.json -o diffamp.sp    # IR → SPICE
layout-canvas schema                                         # Block IR JSON schema
layout-canvas web --port 8080                                # local canvas UI
layout-canvas mcp                                            # MCP stdio server

Web canvas

layout-canvas web opens the editor at http://127.0.0.1:8080: place blocks, edit params, wire nets, then Verify (extract → DRC → LVS, violations painted on the layout) and PEX (post-layout ngspice on extracted netlist with waveform sparklines). Projects save/load as JSON — local-first, nothing leaves the machine.

Python API

from layout_canvas.ir.model import Design
from layout_canvas.tools.verify import verify_design
import layout_canvas.blocks  # registers all block generators

d = Design.model_validate({
    "name": "diffamp", "pdk": "sky130",
    "instances": [{"id": "U1", "block": "sky130.diff_pair",
                   "params": {"fingers": 4, "width": 2.0},
                   "placement": {"x": 0, "y": 0}}],
    "nets": [], "ports": [{"name": "inp", "pin": "U1.inp"}],
})
r = verify_design(d)          # extract → DRC → LVS in one call
r["passed"], r["drc"]["total_violations"], r["lvs"]["match"]

Post-layout simulation on the extracted netlist:

from layout_canvas.tools.sim import simulate_extracted
res = simulate_extracted(d, probes=["out", "tail"], vdd=1.8)
# res["status"], res["waves"]  →  per-probe waveform vectors

MCP (AI agents)

Point any MCP client at layout-canvas mcp (stdio). Agents get the full loop: open_design → insert_block_into_layout / transact → verify_design → run_simulation(source="extracted", probes=[...]) → export_virtuoso / save_project. probe_environment reports which EDA tools are live.

Repository layout

src/layout_canvas/
  ir/          Block IR schema (Pydantic)
  blocks/      parametric generators (sky130/, ihp_sg13g2/)
  compiler/    GDS/SPICE/Skill compile, router, Virtuoso export, PPA
  engine/      transactional edit session (snapshot/undo)
  tools/       drc, extract, lvs, sim, verify, backends
  web/         local canvas app + gallery
  mcp/         stdio MCP server
tests/         245 tests incl. real-ngspice / real-PDK smokes
docs/          architecture, ADRs 0001-0009, external-tool setup
deploy/        Dockerfile + compose for the web app

Security & privacy

• Local-first by design. The web app binds 127.0.0.1; project files, GDS output and simulation decks stay on your machine. No telemetry, no cloud calls, no analytics.
• No secrets in-repo. PDK model paths, tool binaries and simulator paths are supplied via LAYOUT_CANVAS_* environment variables — see docs/EXTERNAL_TOOLS.md. Proprietary PDKs are never shipped or required.
• Fail-closed verification. DRC/LVS/sim report real tool results; missing tools return unavailable/refused, never fabricated green.
• Process isolation. GPL tools (KLayout app, netgen) run as external processes or the pip klayout module — no GPL code is linked in.
• Report vulnerabilities by opening a GitHub issue marked security, or contact the maintainer directly. Please do not include proprietary PDK material or NDA'd data in public reports.

License

MIT — see LICENSE. The "Layout Canvas" name is reserved (see NOTICE); bundled Sky130 model files under examples/models/sky130/ are upstream SkyWater files and remain Apache-2.0. Commercial arrangements (private PDK enablement, enterprise deployment, support): docs/COMMERCIAL.md.

Prior art & references

Independently implemented; product-direction inspiration from Analog Canvas (github.com/cascode-ai/analog-canvas). Built on gdsfactory, KLayout, ngspice, and the Sky130 / IHP SG13G2 open PDKs. Related: Glayout (OpenFASOC), ALIGN, Layout21, BAG, VLSIR.