@casys/mcp-dfm

Measured design-for-manufacturability checks for STEP geometry, exposed as an MCP server. The server snapshots the exact STEP bytes, runs native geometry analysis, and compares the resulting measurements with limits supplied by the caller.

It does not declare a part "manufacturable" or select a printer, material, orientation, or threshold.

Tool	What it measures	Caller-declared comparison
dfm_check_envelope	Axis-aligned X/Y/Z extents, surface-mesh volume status, optional mass	build_volume_mm: { x, y, z } and optional density_kg_m3
dfm_check_overhangs	Downward-facing triangle area and spatial zones	build_direction and max_overhang_deg
dfm_check_min_thickness	Sampled inward-normal wall thickness and thin zones	min_thickness_mm

Every registered check consumes an absolute step_path. Supplying expected_step_sha256 makes the call fail before Gmsh runs if the private input snapshot does not match the expected bytes. Geometry is reported in mm, mm², and mm³; angles are degrees; density is kg/m³; mass is kg.

Measured DFM is not a documentary printability estimate

These tools execute against a STEP snapshot. An empty violations array means only that the measured values did not cross the limits declared for that call. It is not a general printability or supplier-acceptance verdict.

In the Casys Digital Thread integration, the measured industrialize.run-dfm-checks@1 path is deliberately separate from the older documentary industrialize.observe-printability@1 path. The measured case binds an attested STEP, the three-axis build_volume_mm object, and any downstream bed-contact filter. Results from the two paths are not interchangeable.

Mesh evidence and bounded measurement states

Every result includes mesh_topology for the particular Gmsh surface mesh consumed by that call: closed, watertight, manifold, orientation_consistent, connected component count, and the detected boundary, non-manifold, and degenerate counts. These are properties of the emitted triangle mesh, not a claim that the source CAD is valid or that a part is manufacturable. manifold checks both edge uses and each vertex link; watertight is true only when that mesh is closed, manifold, consistently oriented, and has no degenerate triangles.

dfm_check_envelope always returns the numerical divergence-theorem result, but sets measured.volume_status to computed only for one connected watertight shell. Its absolute value is invariant if that one shell is globally reversed, and no second shell exists whose nesting could change the material volume. Multiple closed components remain unverified: this server does not infer their relative orientation or nesting. Derived mass_kg is emitted only when a caller supplies density and that volume status is computed; otherwise mass_status is not_requested or unverified.

dfm_check_min_thickness also returns ray_coverage: the sampled triangle centres, rays that found an opposing surface, unresolved rays, and whether coverage is complete. minimum_thickness_status is unverified unless the mesh is watertight and every sampled ray completed. A sampled result is still an inward-normal sample, not a proof of the global minimum wall thickness.

Quick start: Docker image over HTTP

Version 0.4.1 packages Gmsh, Python, and NumPy for linux/amd64 and linux/arm64. The Docker workflow derives its OCI version from deno.json and accepts a semantic image tag only when the pushed Git tag matches the package version. The command below is the last published 0.4.0 image digest, kept until the 0.4.1 image exists; it is not the 0.4.1 image. After that release workflow completes, pin the new GHCR digest in a deployment manifest.

docker run --rm \
  -p 127.0.0.1:3018:3018 \
  -v /absolute/path/to/step-files:/data:ro \
  ghcr.io/casys-ai/mcp-dfm@sha256:564d3745fd4e0dd35c6eb6a3479292df83362d8a82d00d5812daaff8f23cbe9b http

The image's http mode provides stateless HTTP on http://127.0.0.1:3018/mcp, protocol 2026-07-28. Use paths as seen by the container, such as /data/bracket.step, not host paths. Docker Desktop must be allowed to share the mounted directory where applicable.

HTTP call example

The following source-checkout call measures the committed 40 × 30 × 20 mm fixture. Mcp-Name must mirror params.name:

curl -sS -X POST http://127.0.0.1:3018/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -H 'MCP-Protocol-Version: 2026-07-28' \
  -H 'Mcp-Method: tools/call' \
  -H 'Mcp-Name: dfm_check_envelope' \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
      "name": "dfm_check_envelope",
      "arguments": {
        "step_path": "/data/dfm_healthy_box.step",
        "expected_step_sha256": "0eeade1298710e8ceff66c52f2b7e51bbfb2b3856fafd955fd2cfa6107492693",
        "build_volume_mm": { "x": 200, "y": 200, "z": 200 },
        "mesh_size_mm": 5,
        "density_kg_m3": 2700
      },
      "_meta": {
        "io.modelcontextprotocol/protocolVersion": "2026-07-28",
        "io.modelcontextprotocol/clientCapabilities": {}
      }
    }
  }'

The measured result includes:

{
  "violations": [],
  "measured": {
    "x_mm": 40,
    "y_mm": 30,
    "z_mm": 20,
    "volume_mm3": 23999.999999999967,
    "volume_status": "computed",
    "mass_status": "computed",
    "mass_kg": 0.06479999999999991
  },
  "mesh_topology": {
    "closed": true,
    "watertight": true,
    "manifold": true,
    "orientation_consistent": true,
    "connected_component_count": 1,
    "boundary_edge_count": 0,
    "non_manifold_edge_count": 0,
    "non_manifold_vertex_count": 0,
    "degenerate_triangle_count": 0
  },
  "limits_declared": {
    "build_volume_mm": { "x": 200, "y": 200, "z": 200 },
    "density_kg_m3": 2700
  },
  "input_artifact": {
    "sha256": "0eeade1298710e8ceff66c52f2b7e51bbfb2b3856fafd955fd2cfa6107492693",
    "bytes": 15456,
    "source_path": "/data/dfm_healthy_box.step"
  }
}

Use structuredContent as the machine-readable result. The text content is a short summary for the model. Floating-point values are not rounded in the wire result, so consumers should apply tolerances appropriate to their case.

ghcr.io/casys-ai/mcp-dfm:latest is a mutable convenience tag, not the authority for a version or capability. The digest in the quick-start command is the last published 0.4.0 image. Resolve the 0.4.1 image to its GHCR digest after that image exists.

Tool contracts

Every registered input object is closed (additionalProperties: false). Unknown properties are rejected both by MCP schema validation and when a handler is invoked directly. Numeric sizes and dimensions must be finite and strictly positive: build-volume X/Y/Z, mesh_size_mm, optional cluster_radius_mm, optional density_kg_m3, min_thickness_mm, and optional timeout_ms. sample_count must be a finite positive integer. max_overhang_deg remains 0–90° inclusive; build_direction must be a non-zero three-vector with finite components. These checks run before the STEP snapshot or any native subprocess. The server still does not invent manufacturing defaults or arbitrary maximum caps.

dfm_check_envelope

Gmsh tessellates the STEP into a surface STL. The tool computes the bounding box from vertex extrema, computes volume using the divergence theorem, and computes mass only when density is supplied:

mass_kg = volume_mm3 / 1e9 × density_kg_m3

Field	Required	Description
step_path	yes	Absolute path to the STEP file on the server filesystem
build_volume_mm	yes	Object with strictly positive X, Y, and Z extents in mm: { "x": 200, "y": 200, "z": 200 }
mesh_size_mm	yes	Finite strictly positive Gmsh surface element size in mm
expected_step_sha256	no	Expected 64-character STEP SHA-256; mismatch aborts before Gmsh
density_kg_m3	no	Finite strictly positive caller-owned density used to add mass_kg; no density is inferred
timeout_ms	no	Finite strictly positive Gmsh subprocess timeout in ms; default 60000

The comparison is axis by axis in the STEP coordinate system. The tool does not rotate the part to find a tighter fit. A numerical volume_mm3 is kept for diagnosis, but it is explicitly unverified unless the returned mesh topology establishes one connected watertight shell; derived mass is withheld on that condition.

Measured native fixture result at mesh_size_mm: 5: 40.0 × 30.0 × 20.0 mm and approximately 24000.0 mm³.

dfm_check_overhangs

For each tessellated triangle, the tool measures the angle between its outward normal and -build_direction. The server normalizes any non-zero three-vector. With [0, 0, 1], 0° points straight down, 90° is vertical, and 180° points straight up. A triangle is included when its angle is strictly less than max_overhang_deg; matching triangles are clustered into zones.

Field	Required	Description
step_path	yes	Absolute path to the STEP file
build_direction	yes	Non-zero finite [x, y, z] build direction; [0, 0, 1] is +Z
max_overhang_deg	yes	Finite threshold in degrees from the downward direction, 0–90 inclusive
mesh_size_mm	yes	Finite strictly positive Gmsh surface element size in mm
expected_step_sha256	no	Expected STEP SHA-256
cluster_radius_mm	no	Finite strictly positive spatial merge radius; default 3 × mesh_size_mm
timeout_ms	no	Finite strictly positive Gmsh timeout in ms; default 60000

Important bed-contact behavior: the provider does not remove the bottom face. For the healthy-box fixture at +Z, the returned zone is the 1200 mm² face at Z = -10 mm, with bbox.min[2] and bbox.max[2] both equal to -10. The output does not separately declare the part's global minimum Z, and there is no z_min input. A workflow may remove a bed-contact zone only from an explicit, reviewed plane and tolerance; it must not silently infer one from the lowest reported zone.

The tool also does not model support placement or cost, bridging rules, or material-specific behavior. Curved surfaces are approximated by the mesh.

Measured L-bracket fixture result at mesh_size_mm: 3, 45°, build +Z: overhang area greater than 1000 mm² and at least one zone, including the bed face.

dfm_check_min_thickness

The tool tessellates with Gmsh, samples triangle centres, and invokes Python with NumPy to cast an inward-normal Möller-Trumbore ray. It reports the first opposing surface distance, the minimum sampled distance, sample coverage, and clustered points below the declared threshold.

Field	Required	Description
step_path	yes	Absolute path to the STEP file
min_thickness_mm	yes	Finite strictly positive caller-declared violation threshold
mesh_size_mm	yes	Finite strictly positive Gmsh element size; start at most half the threshold
expected_step_sha256	no	Expected STEP SHA-256
sample_count	no	Finite positive integer triangle centres to sample; default 500
cluster_radius_mm	no	Finite strictly positive spatial merge radius; default 3 × mesh_size_mm
timeout_ms	no	Finite strictly positive total Gmsh and Python timeout in ms; default 120000

This is sampled inward-normal thickness, not a global exact minimum-distance proof. Sampling can miss a small or diagonal thin region. The result reports the exact mesh topology and ray coverage it obtained; minimum_thickness_status is unverified when the mesh is not watertight or any sampled ray did not reach an opposing surface. It does not apply FDM, SLA, SLS, or material-specific feature rules.

Measured thin-wall fixture result at mesh_size_mm: 0.5, sample_count: 300, and threshold 1.0 mm: minimum 0.8000 mm and 12 sampled violations.

Input attestation

Before invoking a native process, each tool:

1. Copies step_path into a private temporary directory.
2. Hashes that private copy with SHA-256.
3. Compares it with expected_step_sha256, when supplied.
4. Makes the snapshot read-only.
5. Passes only the snapshot to Gmsh and returns its digest, byte count, and original source path in input_artifact.

The returned digest therefore identifies the bytes actually consumed. The expectation is optional in the raw MCP schema so exploratory calls are possible; provenance-sensitive workflows should require it.

Run from source or JSR

The HTTP server requires native executables on PATH:

Dependency	Used by	Typical install
gmsh	every registered check	apt install gmsh or brew install gmsh
python3 with numpy	minimum-thickness only	apt install python3 python3-numpy or an activated virtual environment with NumPy

From a checkout:

deno task serve
deno task serve -- --port=3099 --hostname=0.0.0.0

Version 0.4.1 provides both stateless HTTP and native stdio. For stateless HTTP through the JSR package:

deno run -A jsr:@casys/mcp-dfm@0.4.1/server --port=3018

The first two commands expose stateless HTTP.

Native stdio from a checkout, JSR, or published image

Version 0.4.1 provides native stdio. Use an exact entrypoint:

# checkout
deno run -A server.ts --stdio

# JSR 0.4.1
deno run -A jsr:@casys/mcp-dfm@0.4.1/server --stdio

For native stdio from the published image, pass stdio to Docker and keep stdin open with -i:

docker run --rm -i \
  -v /absolute/path/to/step-files:/data:ro \
  ghcr.io/casys-ai/mcp-dfm@sha256:564d3745fd4e0dd35c6eb6a3479292df83362d8a82d00d5812daaff8f23cbe9b stdio

MCP App viewer

Version 0.4.1 includes the MCP App io.casys.mcp-dfm.results@0.4.1. The App is a read-only projection of:

• a closed recorded session io.casys.mcp-dfm.recorded-checks-session/1.0 delivered through viewer.session.apply (the Digital Thread path);
• optional single-tool raw results for the three existing checks, with no invented whole-case verdict.

It does not execute Gmsh, recompute evaluations, or declare a part manufacturable. Recorded Digital Thread evaluations are labelled as Digital Thread-owned. Quality fields that a historical DfmCheckCapture does not record (mesh_topology, volume_status, minimum_thickness_status, ray_coverage) display as unavailable. Private stagedPath / source_path values are stripped or rejected.

The HTML bundle is part of the package layout. Rebuilding it uses the published MCP View source commit b08802df353bb25d25a1c8d64b22ea61b5287ae0, with split packages (@casys/mcp-view@0.9.3, @casys/mcp-view-contracts@0.1.0, @casys/mcp-view-components@0.9.0) and does not fall back to a published SDK:

git clone https://github.com/superWorldSavior/mcp-platform.git .build/mcp-server
git -C .build/mcp-server checkout b08802df353bb25d25a1c8d64b22ea61b5287ae0
export MCP_VIEW_LOCAL_ROOT="$PWD/.build/mcp-server/packages/view"
export MCP_VIEW_CONTRACTS_LOCAL_ROOT="$PWD/.build/mcp-server/packages/view-contracts"
export MCP_VIEW_COMPONENTS_LOCAL_ROOT="$PWD/.build/mcp-server/packages/view-components"
deno task build:ui
deno task check:ui:bundle

The package and App identities are versioned together. The Docker quick-start above retains the previously published 0.4.0 image digest; resolve the new 0.4.1 image after its release workflow succeeds. Consumer runtime qualification and adoption remain separate from publishing this provider release.

Development

deno task release:check
DFM_RUN_NATIVE=1 deno task test

release:check runs formatting, type checking, linting, non-native tests, the stdio wire tests, viewer model tests, and the versioned UI bundle freshness gate. The reusable CI quality workflow runs that fast job without DFM_RUN_NATIVE; its native job installs Gmsh, Python, and NumPy, then runs DFM_RUN_NATIVE=1 deno task test. A pull-request workflow calls the same quality workflow, which must complete before JSR publication or GHCR image build and push.

A workflow publishes a new JSR version only when the version in deno.json is not already present. A separate workflow publishes the multi-arch GHCR image; a semantic image tag is accepted only when it matches that package version.

Security

This server invokes native parsers on caller-supplied files. Keep HTTP bound to loopback unless it is protected by an appropriate trusted boundary. See SECURITY.md for private vulnerability reporting.

License

MIT