DWG MCP Server

DWG MCP Server

DWG and DXF file format access for AI via MCP.

Connect DWG and DXF files to Claude, ChatGPT, Codex, Cursor, and other MCP clients. DWG MCP Server provides structured access to DWG data so AI can understand and reason about drawings. Source files remain read-only; supported entity properties can be changed in memory for inspection and rendering.

Quick Start

Use DWG MCP Server from the MCP client of your choice. The npm package downloads the matching native build from GitHub Releases on first use, verifies its SHA-256 checksum, caches it, and runs it directly. Docker and a separate Python installation are not required.

Native installers

macOS Apple Silicon:

curl -fsSL https://raw.githubusercontent.com/dimitrovakulenko/dwg-mcp-server/main/scripts/install.sh | sh

Windows x64 (PowerShell):

irm https://raw.githubusercontent.com/dimitrovakulenko/dwg-mcp-server/main/scripts/install.ps1 | iex

Both installers download the latest native release and verify its SHA-256 checksum before installation.

After installing, point your MCP client at the installed executable. On macOS add ~/.local/bin to PATH first, or use the absolute path reported by the installer:

claude mcp add --scope user --transport stdio dwg-mcp -- dwg-mcp-server

If your MCP client, AI agent, or test harness does not support MCP roots, set DWG_MCP_ALLOWED_ROOTS to the folders that should be accessible.

Codex

codex mcp add dwg-mcp \

--env DWG_MCP_ALLOWED_ROOTS="$HOME" \ -- npx -y @dmytro-prototypes/dwg-mcp-server

Codex does not currently provide MCP client roots to stdio MCP servers. Configure explicit allowed roots for the folders that contain your drawings.

Claude

claude mcp add --scope user --transport stdio dwg-mcp -- npx -y @dmytro-prototypes/dwg-mcp-server

Claude Code provides MCP roots, so no extra path variable is usually needed.

Cursor

{

"mcpServers": { "dwg-mcp": { "command": "npx", "args": ["-y", "@dmytro-prototypes/dwg-mcp-server"] } } }

Exposed Tools

A typical flow is:

1. Discover available folders with dwg.list_roots.
2. Open a file from one of those folders with dwg.open_file.
3. Inspect supported or file-local types with dwg.list_types, dwg.list_file_types, or dwg.describe_type.
4. Fetch known handles with dwg.get_objects or search the drawing with dwg.query_objects.
5. Optionally change properties reported as writable with dwg.set_entity_properties.
6. Discover and render drawing views with dwg.list_render_views and dwg.render_view.
7. Close the session with dwg.close_file, discarding in-memory changes.

Architecture

Runtime model

In the packaged deployment, DWG MCP Server is a standalone native application built for the host operating system. The Python host exposes the MCP tools, validates file access, and manages document sessions.

Each dwg.open_file call starts a dedicated Rust dwg-worker process for that DWG and returns a host-side documentId. All later file-scoped calls use that document id. dwg.close_file terminates the worker for that session.

Worker and query model

The Rust worker speaks newline-delimited JSON over stdin and stdout. When it opens a DWG through LibreDWG, it first builds an in-memory indexed document. That upfront indexing step is central to the design: the server pays the cost once when the file is opened, then answers later requests against the index instead of rescanning the DWG each time.

The indexed model stores object handles, kinds, type names, generic types, summary and full properties, and derived block, layout, and space membership. It also stores supported type metadata such as aliases, default projections, and property definitions.

When you request full object records, responses also include that derived membership under extendedData, including container block, layout, and model or paper space when known.

dwg.get_objects is direct lookup by handle. dwg.query_objects runs over indices for handle, type, generic type, kind, exact property values, block, layout, and space, then applies filters, scopes, relation traversal, sorting, projection, and pagination. This is what makes queries over blocks, layers, layouts, references, and related objects practical on an opened drawing.

dwg.set_entity_properties validates writes against the type catalog, updates the native LibreDWG document, rebuilds the index, and invalidates the cached rendering scene. Block-reference insertion points use OCS coordinates, block scaling rules are enforced, and attributed block references are rejected until their owned attributes can be transformed safely. The tool does not save or overwrite the source file.

The worker lazily compiles a separate rendering scene on the first render call. It expands block references and generated dimension blocks, draws common 2D geometry, text, MTEXT, hatch contours, and approximated point-list entities, and composes paper-space layouts through their viewports. SVG is the canonical render output; PNG is rasterized from the same SVG in memory. Render responses include coverage diagnostics for generated-block fallbacks and unsupported entity types. Automatic fit rejects extreme sparse extents, WIPEOUT boundaries mask earlier geometry, and explicit regions allow bounded rendering of dense views. See docs/rendering.md for the protocol and rendering model.

Access and packaging

DWG and DXF files must be opened from roots listed by dwg.list_roots. The server first asks the MCP client for roots. If the client, AI agent, or test harness does not support MCP roots, configure explicit allowed roots:

python3 -m dwg_mcp_server --allowed-root "$HOME/Downloads"

or:

DWG_MCP_ALLOWED_ROOTS="$HOME/Downloads;$HOME/Documents/dwg" \

python3 -m dwg_mcp_server

DWG_MCP_ALLOWED_ROOTS is an authorization fallback for clients without roots. It should be a semicolon-separated list of absolute directories.

Official releases contain a standalone MCP host, the statically linked Rust worker, and the LibreDWG schema files needed at runtime.

Native platforms

GitHub Actions builds and tests these release artifacts on their native runners:

x64, x86_64, and AMD64 name the same CPU architecture; the Windows and Linux artifacts work on both Intel and AMD processors.

Build and Test From Source

Local source builds use the vendored third_party/libredwg submodule by default.

Prerequisites

• Rust toolchain
• Python 3.11 or newer
• autotools for local LibreDWG builds on macOS or Linux (autoreconf, aclocal, automake, autoconf, make)

Bootstrap

git submodule update --init --recursive

bash scripts/build-libredwg.sh

Build and test

cargo test --workspace

bash scripts/run-e2e-tests.sh

Run the MCP host locally

The Python host looks for dwg-worker under target/release or target/debug. If you want a release build explicitly:

cargo build -p dwg-worker --release

Then run the MCP host:

PYTHONPATH=server/src python3 -m dwg_mcp_server

If the worker binary lives somewhere else, set DWG_WORKER_BIN to that executable.

Publish a native release

1. Set the same version in npm/package.json, server.json, and the Claude extension manifest when applicable.
2. Configure npm trusted publishing for .github/workflows/native-release.yml.
3. Push the matching v tag.

The Native Release workflow builds and tests all four platforms, creates the GitHub Release with checksum files, and then publishes the npm launcher. A manual workflow run can build one selected platform or all four as downloadable Actions artifacts without publishing.

Clean rebuild

Remove local Rust and Python build artifacts:

bash scripts/clean-build-artifacts.sh

To also wipe the host LibreDWG build under third_party/libredwg:

bash scripts/clean-build-artifacts.sh --with-libredwg

Official MCP Registry

Registry metadata lives in server.json under the GitHub-authenticated name io.github.dimitrovakulenko/dwg-mcp-server.

Before publishing a new registry version:

1. Publish the matching npm package version from npm/.
2. Confirm npm/package.json contains the same mcpName as server.json.
3. Run the manual Publish MCP Registry GitHub Actions workflow, or run:

mcp-publisher login github

mcp-publisher publish

The registry validates the published npm package, so server.json must point to an npm version that already exists on the public npm registry.

License

This project is licensed under the GNU General Public License v3.0. See LICENSE for the full license text.