AutoCAD MCP Pro

Production-grade AutoCAD automation for AI agents — live through COM on Windows, or headless through ezdxf on any platform.

One typed MCP contract controls two execution engines. Build and edit drawings, query exact geometry, apply engineering standards (ISO 128/129/286/1101), refine quality issues inside transactions, and deliver hashed artifacts with validation evidence.

v1.4 release snapshot: 131 tools · 6 resources · 5 prompt templates · 474 collected tests. Runtime discovery through system_about is authoritative.

Install in 10 seconds

pip install autocad-mcp-pro     # or: uvx autocad-mcp-pro
autocad-mcp                     # stdio MCP server, backend auto-selected

Headless (no AutoCAD required, any OS) or live AutoCAD on Windows:

AUTOCAD_MCP_BACKEND=ezdxf autocad-mcp    # portable DXF engine
AUTOCAD_MCP_BACKEND=com   autocad-mcp    # live AutoCAD (pip install "autocad-mcp-pro[com]")

Extras: [com] live AutoCAD (pywin32 + Pillow), [pdf] PDF/screenshot rendering (matplotlib), [full] everything. Working from a checkout is the same commands with pip install -e ".[full]" and python server.py.

Benchmarks

Every number below is generated by a script in benchmarks/ and rendered to SVG with Python (matplotlib). No hand-edited charts, no adjectives.

The leaderboard — nine public AutoCAD MCP servers, one rubric

Fixed public 100-point rubric — CAD breadth (25), correctness and delivery (20), backends/platforms (15), engineering production (15), tests and maintenance (15), security and operations (10) — applied to the public source and documentation of each project (benchmarks/source_review.json, dated). Stars and raw tool counts score nothing. Evidence grade A = source plus locally executed evidence, B = deep source review without a full live run, C = documented tool surface only. This is a source review, not a shared live run — the lanes below are the runtime evidence.

python -m benchmarks.render_chart

1 · Live-run scores — same harness, pinned competitors

Ten fixed tasks, one runner, headless ezdxf lane. Competitors are driven black-box over MCP stdio using their own documented tool contracts at the pinned commit, and every geometry claim is verified by re-opening the exported DXF with ezdxf inside the harness — no self-reporting. unsupported scores zero in the fixed matrix; per-task reasons are committed under benchmarks/results/published/. File-IPC (live AutoCAD) lanes run locally, not in CI.

python -m benchmarks.run_competitors --server puran-water-autocad-mcp --backend ezdxf
python -m benchmarks.run_competitors --server beiming183-autocad-mcp --backend ezdxf
python -m benchmarks.render_live_chart

2 · Task matrix — where the score comes from

The single score hides which capabilities exist, so the matrix shows every task against every server: green = verified pass, amber = partial, gray = the server documents no equivalent (planning, refinement, TABLE/MLEADER, hashed delivery), red = attempted but failed verification.

python -m benchmarks.render_matrix_chart

3 · Headless performance — scale evidence

Workloads call the same backend methods the MCP tools call, so server-side overhead is included. Self-measurement only — competitor servers would pay an extra stdio serialization cost that in-process runs do not, so no cross-server timing claims are made. Numbers move with hardware; the machine fingerprint is recorded in the report.

python -m benchmarks.perf_suite --out benchmarks/results/published/perf-ezdxf.json
python -m benchmarks.render_perf_chart

4 · Version A/B — every release re-proves correctness

compare_versions.py runs the same 21 deterministic, headless correctness checks against the previous release tag and the current tree — each check in its own subprocess, so a hard crash counts as a miss instead of killing the run.

v1.4.0 release gate (baseline v1.3.0, report: ab-v1.3.0-vs-v1.4.0.json):

v1.4.0 added CI, packaging, benchmarks, tool profiles, paper space, ISO 286 fits and opt-in 3D solids without breaking a single correctness check. For contrast, the same suite caught the v1.0.0 → v1.1.0 jump:

python benchmarks/compare_versions.py v1.3.0 --json ab.json

Full rubric, caveats, method and boundaries for every lane: benchmarks/README.md.

What ships in v1.4

The tool surface evolves — ask system_about for the live grouped inventory and system_capabilities for per-backend support modes instead of copying lists from this README.

Architecture

flowchart LR
  A[MCP host / AI agent] --> B[FastMCP server]
  B --> C[Error - audit - timing - logging middleware]
  C --> D[Typed tool contract + tool profiles]
  D --> E[COM backend]
  D --> F[ezdxf backend]
  E --> G[Live AutoCAD]
  F --> H[Headless DXF]
  E --> I[Engineering quality + delivery]
  F --> I

• server.py — FastMCP surface, lifespan, middleware, resources, prompts.
• backends/base.py — shared typed contract + capability model.
• backends/com_backend.py — single-STA-thread COM executor with per-call timeouts.
• backends/ezdxf_backend.py — asyncio.to_thread-wrapped DXF engine with snapshot transactions.
• engineering/ — standards-aware planning, generators, critique, scoring, fits, refinement, validation, delivery.
• security.py — path validation and command/AutoLISP sanitization.

Backend capabilities

system_capabilities returns this machine-readably (native, composite, rendered, snapshot, shared, unsupported) — use it at runtime instead of assuming.

A production drawing workflow

drawing_preflight → drawing_plan → drawing_apply_iso_layers
→ deterministic create/edit tools → dimension_auto (+ fit="H7" callouts)
→ drawing_critique → drawing_refine
→ layout_create + viewport_create → drawing_finalize → drawing_deliver

Connect an MCP client

Claude Desktop / Cursor / any stdio host

{
  "mcpServers": {
    "autocad": {
      "command": "autocad-mcp",
      "env": {
        "AUTOCAD_MCP_BACKEND": "auto",
        "ALLOWED_PATHS": "C:\\Users\\you\\Documents\\AutoCAD",
        "TOOL_PROFILE": "full"
      }
    }
  }
}

After pip install autocad-mcp-pro the autocad-mcp command is on PATH; from a checkout use "command": "python", "args": ["path/to/server.py"]. The model vendor is not part of the server contract.

Local HTTP

autocad-mcp --transport http --port 8000

Loopback only by default; a non-loopback bind is refused unless remote HTTP is explicitly enabled and a bearer token is configured.

Configuration

Security model

Tool input is treated as untrusted:

• every path passes traversal + allowed-root validation;
• dangerous AutoCAD commands and AutoLISP channels are blocked by default, with regression tests for known bypass patterns;
• remote HTTP requires explicit opt-in plus a bearer token;
• DXF size, response sizes, undo history and COM call duration are bounded;
• audit middleware records tool timing and failures;
• system_status / system_about expose unsafe_mode when dangerous operations are enabled.

Please report vulnerabilities privately via GitHub rather than public issues.

Development

pip install -e ".[full]"
pip install pytest pytest-asyncio pytest-cov ruff build
python -m pytest              # 474 tests
python -m ruff check . && python -m ruff format --check .
python -m build

CI runs the same gates on Linux (3.11/3.12), Windows (mocked-COM suite), plus package, Docker and MCP-registry-schema jobs. Releases are tag-driven: git tag vX.Y.Z && git push origin vX.Y.Z builds, publishes to PyPI and cuts the GitHub Release. See docs/RELEASE-DISTRIBUTION.md.

Repository map:

server.py        FastMCP surface and orchestration
config.py        Environment-driven settings
security.py      Path, command, and AutoLISP guards
version.py       Canonical package version
backends/        COM and ezdxf implementations
engineering/     Standards, generators, critique, fits, delivery
benchmarks/      Runners, adapters, published reports, chart renderers
tests/           Headless, mocked-COM, contract, and security tests

Roadmap (1.5)

Screenshot overlay + handle grounding, ezdxf redo (forward snapshots), COM block_create_from_entities, ISO 286 transition/interference hole letters (delta rule), titleblock on paper-space layouts. Features are published when their contracts and limitations are testable — not when they make a longer checklist.

Contributing

Pull requests are welcome — especially reproducible competitor adapters, mocked/live COM coverage, engineering standards, and backend parity. For a large change, open an issue first so the tool contract and evidence plan can be agreed before implementation.

Author

Umutcan Edizsalan · Mechanical engineering work at Anka-Makine · GitHub @U-C4N

Built from production drawing work, then made model-agnostic through MCP.

License

MIT

mcp-name: io.github.u-c4n/autocad-mcp