Forge-Mind / CadKitty

Editable CAD tools for AI assistants, with saved revisions and measured evidence.

Forge-Mind lets an assistant build parts and assemblies, inspect saved geometry, revise a candidate, and export CAD files. CadKitty is the workspace and MCP tool interface. The core is Python; this snapshot does not include the hosted website or deployment setup.

A 42-component tool fixture, exercised through the real MCP interface. It includes editable source, downloaded CAD, measured jaw spacing and a discrete motion check. It is a scripted tool example.

An actual generated inline-four study, shown with the block and covers hidden. The study remains incomplete. The example includes exact STEP files, instructions, source and unresolved work.

What it does

• Build editable native features and assemblies, then change named features or parameters.

• Read saved source, topology, sections, dimensions, and scoped engineering results.

• Solve coupled planar mechanisms in XY, YZ, or ZX, with sampled motion and interference checks.

• Retain durable revisions, failed attempts, measurements, and artifact hashes for review and repair.

• Export saved STEP, STL, and GLB artifacts through the supported tool routes.

flowchart LR
    A[Request and sources] --> B[Authored CAD]
    B --> C[Saved revision]
    C --> D[Geometry inspection]
    D --> E[Review and export]

Each check has its own scope. CAD validity and renders do not certify manufactured fit, strength, safety, or complete engineering function.

Related MCP server: gitcad

Start with a deterministic example

Install uv and use Python 3.11. These commands use Windows PowerShell, from the repository root. The project declares Python 3.11 and 3.12 support; consult the release qualification for tested platforms.

uv sync --locked --python 3.11
uv run --locked forge-mind version
uv run --locked forge-mind validate tests/fixtures/handwritten/nema17_mount_good.partdsl.json --out-dir ./local-output/nema17

The last command compiles a hand-authored reference and runs its validators. It does not call a model. Read each check and its scope; the command's overall result is not physical qualification.

Connect a local assistant with MCP

Use a private local workspace for trusted inputs. Install the workspace extras:

uv sync --locked --extra server --extra mcp --extra fea --python 3.11

Start the service in a separate PowerShell terminal:

$env:FORGE_WORKSPACE_LOCAL='1'
$env:FORGE_WORKSPACE_EXECUTOR='trusted_local'
$env:FORGE_WORKSPACE_DIR="$env:LOCALAPPDATA/ForgeMind/cad-tools"
$env:FORGE_FORBID_API_SPEND='1'
uv run --locked --extra server --extra mcp --extra fea uvicorn forge_mind.server.workspace_app:app --host 127.0.0.1 --port 4321

Keep this service on loopback. It can execute authored CAD on your computer. Choose a new workspace directory for experiments.

Add this server to your MCP-capable assistant. Replace the repository path:

{
  "mcpServers": {
    "cadkitty": {
      "command": "uv",
      "args": ["--directory", "C:/path/to/Forge-Mind", "run", "--locked", "--extra", "server", "--extra", "mcp", "--extra", "fea", "cadkitty-mcp"],
      "env": {"FORGE_MCP_WORKSPACE_URL": "http://127.0.0.1:4321"}
    }
  }
}

Keep the same extras in each command. The assistant starts the stdio bridge; the bridge alone waits for a client. Set the same loopback URL on both sides if you change the port.

Ask the assistant to call get_status first. It reports the workspace connection and current capabilities. Exports default to ~/Downloads/CadKitty; set FORGE_MCP_EXPORT_DIR in the MCP server environment to choose another directory.

With ordinary CAD tools, the calling assistant authors the design. create_project creates an empty project; submit_document queues CAD without a built-in model call. The ai_create_project and ai_edit_project tools use the workspace's installed, signed-in Codex CLI. Set FORGE_CODEX_EXE if that executable is outside PATH. Built-in AI calls can use subscription limits or credits; FORGE_FORBID_API_SPEND=1 does not make them free.

Examples and evidence

See EXAMPLES.md for deterministic references and their limits. Architecture explains the code paths; principal-plane mechanisms describes the coordinate rules.

SNAPSHOT.json records source origin, file hashes, and exclusions. RELEASE_QUALIFICATION.json records the standalone checks and exact tested bytes. Read evidence for the exact release instead of treating historical test totals as current results. Historical north-star specifications have separate availability notes.

For development, see CONTRIBUTING.md. For trust boundaries and reporting, see SECURITY.md.

Read the reproduction guide, mechanism coordinate contract, and measured engine comparison for details.

License

MIT, copyright Brennen Johnson. Third-party packages and imported assets retain their own terms. Optional extras can add licensing and platform requirements.