What are the best open source MCP server CAD tools in 2026, and can AI actually control CAD software?

Yes — open source MCP (Model Context Protocol) servers now let AI agents like Claude, ChatGPT, and GitHub Copilot directly create and modify CAD models, and by September 2026 the ecosystem has matured far beyond novelty demos. The most practical entry points are FreeCAD MCP servers (fully open source, free), PTC's Onshape FeatureScript MCP server (free for hobbyist/maker accounts), Bentley's STAAD MCP server (published to the open MCP Registry in May 2026), and community projects like the Rust-based Autodesk Platform Services CLI with MCP support. This article breaks down what actually works, what is still hype, and how to decide which route fits your workflow.

The Direct Answer: Which Open Source MCP CAD Tools Exist Right Now

Also worth reading: How Does Automated Building Code Compliance Software Actually Function in Modern Architectural Workflows? · What Is the Best Drawing Review Software for Architects in 2026? · How Does Architectural Plan Review Automation Actually Work in 2026?

As of 5 September 2026, the open source MCP server CAD ecosystem clusters into four groups. First, FreeCAD MCP servers: FreeCAD is GPL-licensed and scriptable in Python, which makes it the easiest target for AI control. Multiple community-built MCP interfaces appeared through 2025 and were covered by outlets like 3Druck.com, which demonstrated Claude and ChatGPT controlling FreeCAD models through natural-language prompts. These projects are genuinely open source — you can read the code, fork it, and audit exactly what the AI is allowed to execute.

Second, vendor-published servers with open protocols. PTC shipped an Onshape FeatureScript MCP server that converts text prompts into FeatureScript code executed in Onshape's cloud CAD — described by Embedded Computing Design as a text-to-code-to-CAD pipeline. Onshape itself is proprietary SaaS, but the MCP server layer and FeatureScript language are documented and inspectable, and Onshape's free tier makes it accessible to individual engineers. Bentley Systems published its STAAD MCP server to the open Model Context Protocol Registry in May 2026, connecting AI agents directly to structural analysis workflows — a notable move because STAAD sits in the AEC (architecture, engineering, construction) space rather than mechanical CAD.

Third, CLI wrappers and aggregation layers. 'Raps', a Rust CLI for Autodesk Platform Services with zero-millisecond cold-start dependencies and MCP support, appeared on Hacker News as a Show HN launch, letting AI agents query and manipulate APS (the platform behind Fusion and AutoCAD cloud services) without heavyweight SDKs. Meanwhile projects like One-MCP tackle the context-bloat problem: a single MCP server that exposes thousands of tools without stuffing every tool schema into the model's context window.

Fourth, there is a darker corner: security researchers documented that MCP servers were mass-forked and republished on public registries, creating a supply-chain attack vector. If you install an MCP server that can execute code against your CAD files or your Autodesk account, installing a typosquatted fork is a genuine risk. Always verify the publisher, star history, and commit lineage before connecting a server to anything with credentials.

How MCP Actually Connects AI to CAD Software

The Model Context Protocol, originally open-sourced by Anthropic in late 2024, standardizes how an AI client (Claude Desktop, Claude Code, IDEs like Cursor, or agents such as GitHub Copilot CLI and OpenAI Codex) talks to external tools. An MCP server exposes a list of functions — create_sketch, extrude, boolean_cut, export_step — with typed parameters. When you ask Claude to 'model a bracket with two 6mm mounting holes', the model selects the right tool calls, fills in parameters, and the server executes them against the CAD kernel or API.

Three architectural patterns dominate. Direct kernel scripting: the MCP server wraps the CAD application's scripting interface (Python for FreeCAD, FeatureScript for Onshape, the Revit API for Revit). This gives the AI full parametric control but requires the application to be running or headless-accessible. Cloud API wrapping: the server calls REST APIs like Autodesk Platform Services or Onshape's API — no local installation needed, but you are rate-limited and dependent on vendor uptime. File-based intermediation: the server reads and writes neutral formats (STEP, DXF, IFC) and the AI reasons over geometry data. This is the pattern most relevant to architecture, where drawings and models live in DWG, DXF, RVT, or IFC files that need parsing before any code conversion happens.

The iteration loop matters more than the first prompt. A competent setup lets the agent inspect the result — query vertex coordinates, check bounding boxes, run a mesh validity check — and correct itself. Without that feedback loop, AI-generated CAD degenerates into guesswork. The FreeCAD MCP implementations that work well in practice expose query tools alongside mutation tools; the ones that only write geometry produce models that look plausible and fail on manufacturability.

FreeCAD: The Purest Open Source Option

FreeCAD deserves its own section because it is the only fully open source CAD package with meaningful MCP traction. The application's Python console exposes nearly the entire feature tree, so an MCP server is essentially a JSON-to-Python bridge. Community setups reported on 3Druck.com showed ChatGPT and Claude building parametric parts, adjusting dimensions via follow-up prompts, and exporting STEP files — all without the user touching the GUI.

The strengths are real: zero license cost, no cloud dependency, complete auditability of the server code, and a large Part/PartDesign API surface. The weaknesses are equally real. FreeCAD's kernel (OpenCASCADE) occasionally produces invalid geometry from aggressive boolean operations, and an AI agent does not always notice. Performance degrades on assemblies beyond a few hundred parts. And the community MCP servers vary wildly in quality — some are weekend projects with three tools, others expose dozens of well-typed functions. Expect to spend an afternoon evaluating two or three forks before finding one that handles your use case. For hobbyists, educators, and small shops, that trade is usually worth it. For teams with contractual requirements around model provenance, the maturity gap versus vendor-supported servers is a legitimate reason to wait.

Comparison: Open Source MCP CAD Options Side by Side

Two observations cut against the hype. First, 'MCP support' is becoming a marketing checkbox — a server that exposes two read-only tools technically qualifies but delivers little. Second, the aggregation projects (One-MCP, Strata from YC P25's batch, which pitched one MCP server handling thousands of tools) solve a genuine scaling problem but add an abstraction layer that can obscure errors. If a tool call fails inside an aggregator, debugging which underlying server misbehaved costs time.

Architecture and AEC: Where This Meets Building Design

For readers in architecture rather than mechanical design, the landscape differs in an important way: the artifacts are drawings, BIM models, and code conversions, not machined parts. Bentley's STAAD MCP server signals that structural workflows are the first AEC beachhead, because structural analysis is inherently parameter-driven and thus well suited to AI agents. Architosh's coverage of Autodesk Revit 2027 highlighted big AI and graphics changes, though Autodesk has kept its MCP strategy more guarded than Bentley or PTC — most Autodesk connectivity today flows through community projects like Raps rather than first-party servers.

This is where drawing-to-code conversion becomes practical. An MCP-enabled pipeline can ingest a floor plan (DXF, PDF, or raster), extract wall segments, doors, and dimensions, and emit structured output — JSON geometries, IFC entities, or executable scripts. The reliability ceiling is real, though: raster plans with hand annotations still require human verification, and dimension inference from poorly drafted drawings produces errors that propagate silently downstream. Rule of thumb based on current tooling: vector-sourced plans convert with roughly 90–95% geometry accuracy, while scanned legacy drawings drop to 60–80% and always need review. Treat AI output as a first draft that eliminates 70% of manual tracing work, not as a finished deliverable.

Common Mistakes When Adopting MCP CAD Servers

The most frequent error is connecting an MCP server with write access to production file storage without sandboxing. An AI agent instructed to 'clean up the model' can delete feature history or overwrite assemblies in ways that are expensive to reverse. Run against copies or version-controlled repositories first — both FreeCAD's native format and STEP exports diff poorly, so use Git on FCStd files (they are zip archives and diff poorly, but history still protects you) or rely on Onshape's built-in versioning.

The second mistake is ignoring supply-chain hygiene. The documented mass-forking and republication of MCP servers means the version you install from a registry may not match the upstream repository. Pin versions, verify checksums where provided, and prefer servers whose maintainers publish under verifiable organizational identities. The third mistake is expecting text-to-CAD to replace modeling skill. The MakeUseOf author who let Claude design 3D parts concluded it works well for well-specified, moderately complex geometry — brackets, enclosures, fixtures — and poorly for organic surfacing or tightly toleranced mechanisms. AI accelerates competent modelers; it does not substitute for understanding draft angles, tolerances, and manufacturing constraints.

Finally, people confuse context-window budgets. Every tool exposed by an MCP server consumes tokens describing its schema. Loading a server with 200 tools can crowd out the actual reasoning, which is precisely the problem One-MCP and Strata target. Keep your tool set minimal and task-scoped.

Costs, Licensing, and When to Adopt

Pure open source routes cost nothing in licensing: FreeCAD and its MCP servers are free, Raps is free (though Autodesk Platform Services API usage is metered), and Onshape's free tier suffices for evaluation, with paid plans typically starting around $1,500 per user per year for professional use. Bentley STAAD requires a commercial license, so its MCP server only matters if you already own STAAD or plan to. The hidden cost is time: budget 4–12 hours to stand up a working local setup (install FreeCAD, install an MCP server, configure Claude Desktop or Claude Code, debug the connection) and another few hours learning prompt patterns that produce valid geometry.

On timing: if you are a hobbyist or educator, adopt now — the FreeCAD ecosystem is already productive. If you are a product engineering team, pilot Onshape's FeatureScript server on non-critical parts this quarter; vendor support lowers operational risk. If you are in AEC and need structural analysis integration, Bentley's May 2026 registry publication makes STAAD worth a pilot if licensed. If your goal is converting existing architectural drawings into structured code or models, dedicated conversion platforms that handle drawing parsing end-to-end will outperform DIY MCP chains for the foreseeable future, because the hard part is robust geometry extraction from messy drawings, not the AI-to-CAD interface itself.

The pragmatic 2026 position: open source MCP CAD tools are production-usable for mechanical prototyping and structural workflows, still immature for fully automated architectural deliverables, and improving fast enough that whatever you build should be designed around swappable server components rather than a single vendor's stack.