Drawarch

What drawarch MCP server does

The drawarch MCP server turns architecture descriptions and reference images into portable .drawio artifacts. Generated diagrams use native Draw.io objects rather than a pasted screenshot, so containers, labels, icons, shapes, connectors, ports, and waypoints remain independently editable. Embedded assets keep the resulting file usable offline.

For prompt-based generation, the server can apply one of six built-in themes, search its bundled icon catalog, resolve approved assets, create a deterministic diagram, and validate the resulting file. For image recreation, it accepts an absolute scene graph supplied by the MCP host and preserves geometry, layers, z-order, grouping, labels, and connector routing where represented in that plan.

How it works

The MCP host performs the visual reasoning. DrawArch does not call another language model and does not require a second model API key. A host with image understanding can inspect a reference image and submit a typed scene graph to the server.

Reference recreation follows a reviewable revision flow:

1. prepare_reference_recreation stores the editable scene graph.
2. The user or host policy reviews the plan and any online asset sources.
3. update_reference_plan creates a new immutable revision when changes are needed; earlier approvals become invalid.
4. create_reference_drawio renders an approved revision, embeds assets, validates the file, and returns the artifact.
5. compare_reference_recreation reports structural coverage and missing editable elements for further iteration.

This process targets high-fidelity editable reconstruction, not a guaranteed pixel-identical copy.

Setup and configuration

Node.js 20 or newer is supported, and a Docker image is available. The published package can run through npx and uses stdio by default. Diagnostics are written to stderr so stdout remains available for MCP messages. A source installation uses npm ci, followed by the project test and build commands.

The server can also run as Streamable HTTP in the published container at /mcp, with a public /health endpoint. HTTP MCP access and file downloads require bearer authentication. Container deployments can configure an output volume, API key, approval secret, allowed hosts, and allowed origins. Remote deployments need operators to provide the HTTPS domain and infrastructure; no managed hosted endpoint is provided.

Tools and capabilities

• list_themes returns the six built-in visual themes.
• list_assets searches the bundled real-world icon catalog.
• resolve_asset resolves bundled icons or opt-in online SVGs and reports provenance.
• create_drawio produces a deterministic themed architecture diagram.
• validate_drawio checks a generated architecture file.
• prepare_reference_recreation, update_reference_plan, and get_reference_plan manage reviewable scene-graph revisions.
• create_reference_drawio renders an approved reference plan into a portable file.
• validate_reference_drawio checks native structure, asset references, and embedded-image safety.
• compare_reference_recreation measures structural coverage and identifies missing editable elements.

Limitations and notes

The host must supply the visual analysis for reference images. DrawArch renders and validates that interpretation rather than independently understanding the image. Online assets are opt-in and can be subject to approval. The output is designed for editable structure and controlled iteration, not pixel identity. ChatGPT cannot start a local stdio process, so it requires a remote Streamable HTTP deployment behind HTTPS.

The drawarch MCP server supports local stdio and remote Streamable HTTP operation. For shared services, use authentication, stable approval secrets, exact host and origin allowlists, TLS, and persistent output storage as appropriate for the deployment.