agents
Your agent already
knows how to use it.
vcad is agent-native: 140 MCP tools over the same kernel humans use, with fail-closed checks no prompt can talk its way past.
watch one work
a real session shape: model → check → quote → ask permission. the agent cannot skip the middle two.
the vocabulary
140 tools. One kernel.
Modeling, inspection, sheet metal, boards, physics, fabrication — every tool returns data an agent can act on, and mutations report exactly what changed.
fail-closed
Trust the work,
not the model.
Agents hallucinate; kernels don't. Every agent edit runs the same DFM, clearance, and DRC checks a human's would — and export is blocked until they pass. The receipt is how you audit what your agent did.
export_gerber()
blocked · drc: 2 clearance violations
fix or ask — the board doesn't ship dirty
connect
Two lines, any agent.
# claude code
claude mcp add vcad -- npx -y @vcad/mcp
# anything that speaks mcp
{ "mcpServers": { "vcad": { "command": "npx", "args": ["-y", "@vcad/mcp"] } } }
claude code · cursor · codex — if it speaks mcp, it has a seat