What file formats does Cad Mcp Server support?
The server works with STEP files (explicitly mentioned as the format it reads, measures, and compares).
@mattmohandiss
基本情報
AI-native CAD inspection, no CAD license required. CAD MCP Server bundles a WebAssembly Open CASCADE kernel so LLM tools can read, measure, and compare STEP files locally. The server returns measured facts; your AI assistant interprets engineering meaning.
以下の設定を使って、このサーバーを MCP 対応クライアントに追加してください。
{
"mcpServers": {
"cad": {
"command": "npx",
"args": [
"-y",
"cad-mcp-server"
]
}
}
}Use this FIRST when given a STEP file to inspect. Returns a compact summary: bounding box (AABB and OBB), watertight status, body/solid/face/edge/vertex counts, global properties (volume, surface area, center of mass, inertia, principal axes), per-subshape validity, tolerance statistics, and XDE metadata (assembly tree, PMI summary, color/layer/material presence). If the part is invalid, returns a structured list of which sub-shapes failed and why — the LLM gets actionable information, not a boolean. Don't use for entity-level searches; use query_step for that.
Declarative query over a STEP file's geometric, topological, and XDE entities. Specify which entity type to query (faces, edges, bodies, vertices, pmi, color, layer, material, assembly_node), filter by properties, optionally group by a shared dimension (axis, normal, surface type, etc.), measure derived values per entity or per group (ray tests, distances, curvature, section, continuity), and aggregate statistics (count, min, max, avg) over the result set. Use group_by to cluster entities that share a property — the operation find_coaxial_cylinders used to do is now {entities: "faces", filter: {surface_type: "cylinder"}, group_by: ["axis"]}. The result is one call instead of the 5+ round-trips that primitive-only tools require, and intermediate state stays on the server (set return_intermediate: true for debugging). For multi-step workflows that need iteration across result sets ("for each hole, ray-test, then filter"), use transact_step instead.
Compare two STEP files and return metric deltas, topology changes, body-level changes, and XDE-level changes (PMI, colors, materials, assembly). Use this to compare two revisions of a part when you need factual differences in dimensions, volume, area, topology counts, or exchange metadata. Does NOT track feature identity across revisions — a hole that moved is "hole removed + hole added," not "hole moved." Deltas are comparison_file_path minus baseline_file_path.
Run a multi-step workflow that needs iteration across result sets — for example, "for each hole, ray-test in +Z, then find ones where the ray didn't come out the other side." Each pipeline is a sequence of typed operations. The vocabulary is small: query (re-uses the query_step shape), for_each (apply a sub-pipeline to each item), filter_results (keep items where a condition holds), select (project to specific fields), and walk_assembly (XDE: traverse the assembly tree). Use this only when a single query_step call cannot express the workflow; for declarative queries, prefer query_step. Intermediate results are hidden by default; set return_intermediate: true to debug.
Cad Mcp Server is an AI-native CAD inspection tool that bundles a WebAssembly Open CASCADE kernel so LLM tools can read, measure, and compare STEP files locally without requiring a CAD license.
—
The server works with STEP files (explicitly mentioned as the format it reads, measures, and compares).
The server bundles a WebAssembly Open CASCADE kernel, so no CAD license or external CAD software is required.
It returns measured facts; the AI assistant interprets the engineering meaning.
The README states it reads, measures, and compares; no mention of editing capabilities is provided.
28-tool MCP server for web scraping, crawling, deep research & autonomous extraction — clean Markdown & structured JSON for Claude, Cursor & any MCP client. 1,000 free one-time credits, local-Ollama LLM support.
TruVerifAI's Panel Review is the guardian agent for AI's highest-stakes coding decisions: four frontier models argue over an AI coding agent's riskiest designs, diffs, and commits before they execute, with local review g
AON MCP connects AI agents to real, buyable products and services with live pricing and tracked links—enabling builders to earn developer revenue share from attributed clicks and conversions.

Quantum-inspired keyring for AI coding agents. Secure secrets with superposition, entanglement, tunneling, and teleportation — CLI + 44 MCP tools.

AI growth agent for software businesses. Ranla ships pages that rank on Google, gets your product named in AI answers, replies on Reddit and sends lifecycle email. Ask how growth is going, have Ranla draft the work and a
コメント