On 29 June 2026 an Autodesk Fusion blog post gave advice to people connecting Claude to Fusion, and it is more candid than most vendor copy. The connector is "not yet reliable for generating complex CAD models or full assemblies", the post says, but "it excels at tasks like reporting, screenshots, bulk edits, and data extraction". This is Autodesk's guidance on its own connector, not the result of a published test.
The measurements come from benchmarks, which score models writing CAD code in a test harness and do not test any MCP server or CAD connector. On BenchCAD 2.0, 100 held-out cases across six tasks including assemblies, the best result its maintainers have run is 0.59 out of 1 (October 2026). In the assembly tasks of three frontier models, 83–96% of the part types that scored zero had the right shape in the wrong place.
This report is built from public sources read up to the research cutoff of 20 November 2026: vendor documentation and release notes, repositories, benchmark papers and practitioner accounts. DataRoot Labs is an AI consultancy that builds AI systems for companies. Statuses are the vendors' own labels with dates, and vendor claims are marked as such.
The short answer, package by package
Agents connected to CAD are useful today for queries, reports, bulk edits and simple parts with every dimension stated, not yet for assemblies or guessed dimensions, so the safest first test is an official MCP server used read-only, or code-CAD on a copy of the design.
Scroll horizontally to explore all columns →
| Package | Built-in AI (status) | Official MCP server (status) | What to try first |
|---|---|---|---|
Fusion | Autodesk Assistant, GA since AU 2026 (Sept 2026); agentic use metered | Fusion MCP and Fusion Data MCP, GA on Autodesk's index; Fusion Compute MCP, Public Beta | Fusion MCP, read-only at first |
Inventor | Assistant acting on the model (iLogic, Python), 2027.2; no GA label | Inventor MCP Server, Public Beta, 2027.2+ | Inventor MCP Server on a test seat |
AutoCAD | Assistant, tech preview: standards checks | AutoCAD and Civil 3D MCP Server, Tech Preview (Public Beta); only client is Autodesk Assistant | A community server in read-only mode |
Revit | Assistant, Tech Preview from April 2026 | Revit MCP Server, Technical Preview, Revit 2027.2 | Revit MCP Read tools |
SOLIDWORKS | Local ML features (released); LEO cloud skills, Beta labels, token-metered | None; Dassault names A2A as its interop route | LEO drawing skills, or a community server on a non-production seat |
CATIA | AURA, LEO, MARIE on 3DEXPERIENCE SaaS (July 2026) | None | Built-in LEO skills |
Onshape | FeatureScript MCP, Onshape Labs early access (Aug 2026) | FeatureScript MCP, watching API quotas | |
Creo | Creo AI Assistant: Advise GA, Assist beta, Automate closed beta | None; three-phase plan, no dates | Assist tier, read-only |
Designcenter (NX) | Copilot (documentation Q&A), Command Prediction | Designcenter MCP announced on a marketing page only | Copilot for documentation questions |
Rhino | — | RhinoAI (McNeel), pre-1.0, no status label | RhinoAI |
FreeCAD | None | None | The community server with an approval dialog for code |
Code-CAD (CadQuery, build123d, OpenSCAD) | n/a | Zoo's official server for its own KCL language | build123d or CadQuery if you need STEP |
Archicad | Archicad MCP Server, Beta, Collaborate/Max editions | Official server (Beta) | |
KiCad / Altium | None / Requirements Portal agents | KiCad none; Altium "Coming Soon" | KiCad community server on a copy of the project |
The last column is an editorial suggestion drawn from the evidence below, not a ranking. Community servers are listed in the MCP section.
Terms used in this report
- MCP (Model Context Protocol): an open protocol that lets Claude and other agents call tools inside an application.
- stdio: the client starts the MCP server as a local child process and talks to it over standard input and output, with no network port.
- Code-CAD: modelling by writing a program (CadQuery, build123d, OpenSCAD, KCL) that a geometry kernel turns into a part.
- B-rep / STEP vs mesh: a B-rep stores exact faces and edges and exports to STEP for machining and drawings; a mesh (STL, OBJ) is a surface of triangles with no exact faces or parameters.
- Feature tree: the ordered sketches and features a parametric CAD rebuilds when a dimension changes.
- IoU (intersection over union): the overlap between a generated solid and the target as a share of their combined volume, from 0 to 1. BenchCAD 2.0 scores a part as 0.8 × 3D overlap plus 0.2 × topology, so its scores are averages, not pass rates.
- F1: a 0–1 score balancing how many expected items a model found against how many of its outputs were correct.
- DNS rebinding: an attack in which a web page tricks the user's browser into sending requests to a server on the user's own machine.
- Flex tokens: Autodesk's consumption currency; in Fusion, one token equals 120 "AI assists".
- Status labels: GA (generally available); Public Beta (open to all, may change); Tech or Technical Preview (early, no commitments); Labs or early access (opt-in); Coming Soon or announced (not usable yet).
Three ways AI gets into CAD
AI reaches CAD through assistants the vendor builds in, plug-ins that do one job, and general agents such as Claude, ChatGPT or Gemini that drive the CAD from outside. Which route reaches your package differs by vendor, and several answers changed in the last six months.
Built-in assistants. The vendor chooses the model, limits what it can do and meters it. Fusion's Autodesk Assistant moved from a help chatbot to text-to-command in March 2026. From October, agentic use draws on a monthly allowance, labelled beta, of 20 Flex tokens for commercial and startup subscriptions and 10 for personal, education and trial users; help questions stay free. SOLIDWORKS splits its AI into local features and cloud "virtual companions" that, in the SOLIDWORKS 2027 FAQ's words, are "measured through a consumption model using Tokens". Siemens' Copilot is mostly documentation Q&A, powered by Microsoft Phi-3. Autodesk builds its assistants on the same tool layer as its MCP servers: the Inventor MCP server has the "same set of tools used by the built-in Autodesk Assistant", and the Revit Assistant is "powered by Model Context Protocol".
Single-job plug-ins. A third party solves one task: CAM programming (CloudNC CAM Assist), drawings (DraftAid), design review (CoLab AutoReview), sketch constraints (Hestus) or scan to feature tree (Backflip). Prices are mostly "on request"; Hestus is free and Werk24 publishes €1.25 per page. Evidence beyond vendor claims is thin for almost all of them.
General agents from outside. You bring the model and pay for its API use, plus any vendor quota; an MCP server, the vendor's API or a code-CAD library gives it hands. Agents break most often on assemblies, placement, inferred dimensions and silent failures. Autodesk, PTC for Onshape, McNeel and Zoo ship official CAD servers, as do Graphisoft and Bentley in AEC. Code-CAD is where progress can be measured: most 2026 CAD benchmarks, including BenchCAD, CADBench and MUSE, ask models to write CadQuery.
What MCP is, and which CAD packages have a server
MCP was released by Anthropic in November 2024 and donated to the Agentic AI Foundation, a directed fund under the Linux Foundation, on 9 December 2025. The current revision is dated 2026-07-28 and made the protocol stateless. A host application such as Claude Desktop, Cursor or VS Code runs a client, the client talks JSON-RPC to one or more servers, and each server offers tools, resources and prompts.
For CAD, a server is a translator between tool calls and the CAD's own programming interface. Autodesk says the Fusion connector "gives Claude access to the same Fusion API that scripts and add-ins use". An MCP server therefore adds no modelling capability the API lacks; it decides which parts of the API a model can reach and how results come back. Three designs recur:
- an add-in inside the CAD process that opens a local socket, with a small MCP server outside it (the faust-machines Fusion server, rhinomcp, freecad-mcp);
- COM automation of a running Windows application from an external process (SOLIDWORKS, CATIA V5, Solid Edge);
- a headless kernel or cloud API with no desktop session (Fusion Compute MCP, Zoo, build123d-mcp).
Official and community servers in November 2026
Autodesk gets the most space here because it publishes the most MCP documentation; less documentation from Dassault, PTC and Siemens does not necessarily mean less capability. On 20 November 2026 Autodesk's MCP hub listed six servers: three for Fusion, one for AutoCAD with Civil 3D, one for Revit and one for product help. Inventor 2027.2 documents its own server in a separate help section. Dassault has none for SOLIDWORKS or CATIA, PTC has one narrow server for Onshape, Siemens has announced one, and Solid Edge and Creo have none.
Scroll horizontally to explore all columns →
| Server (maintainer) | Status (date) | Connection and authentication | Runs model-written code? |
|---|---|---|---|
Fusion MCP (Autodesk) | GA on Autodesk's index, no label on its help page; tech preview at announcement, 15 Apr 2026; registry v1.0.1, 29 Oct 2026 | Local HTTP in the desktop client, none | Yes: Fusion API scripts |
Fusion Compute MCP (Autodesk) | Public Beta (Sept 2026) | Cloud, headless, OAuth | Yes: TypeScript |
Inventor MCP Server (Autodesk) | Public Beta, 2027.2+ | Local stdio via named pipes; Autodesk sign-in | Yes: Python, iLogic |
AutoCAD and Civil 3D MCP Server (Autodesk) | Tech Preview (Public Beta) | Local, inside Autodesk Assistant only | No code tool listed |
Revit MCP Server (Autodesk) | Technical Preview, 2027.2 | Local stdio | No code tool in any set |
FeatureScript MCP (PTC) | Labs early access (13 Aug 2026) | Cloud HTTP | FeatureScript in Onshape's cloud |
RhinoAI (McNeel) | No label; v0.4.0 (3 Nov 2026) | Plug-in plus stdio router | Yes: Python, C#, commands |
zoo-mcp (Zoo) | v0.35.1 (20 Nov 2026) | stdio or HTTP; Zoo login or token | KCL on Zoo's hosted engine |
Designcenter MCP (Siemens) | Announced | Not documented | Not documented |
freecad-mcp (community) | Pushed 16 Nov 2026 | Add-on RPC socket | execute_code, on by default |
FreeCAD robust MCP server (community) | Pushed 20 Nov 2026 | JSON-RPC socket | Off by default; approval dialog |
fusion360-mcp-server (faust-machines) | Beta; pushed 6 Oct 2026 | Local TCP, none | execute_code |
SolidworksMCP-python (community) | Registry v2.0.0 (13 Nov 2026) | COM | Macros |
onshape-mcp (community) | Pushed 10 Sept 2026 | REST, API keys | FeatureScript |
build123d-mcp (community) | Pushed 19 Nov 2026 | Local kernel, no CAD seat | Python |
SolidworksMCP-python publishes a support level for each tool, though its README gives three different tool counts.
Fusion. The local server is switched on in Preferences and listens at 127.0.0.1:27182. Its help page names Claude Desktop, Cursor and VS Code as clients, says it is desktop-only, and publishes no tool list: "Available tools are discovered by your MCP client at connection time and may change." The Claude connector built on it was announced on 28 April 2026 "for Fusion subscribers". No Autodesk page found for this report says that MCP use consumes Flex tokens.
Fusion Compute, the headless option, allows two concurrent sessions per user and stops a session after 15 minutes idle, and needs an active Fusion licence plus a Fusion Team hub.
AutoCAD and Inventor. AutoCAD does have an official server, but its documentation names Autodesk Assistant as the only supported client and its tools read and analyse; the November 2026 release notes add property editing. Inventor's server is meant for outside agents: its page names "Claude Desktop, Claude Code, Cursor, and other AI clients that support local MCP servers", allows one agent at a time and lists 46 tools, including execute_python_script and execute_ilogic_code.
Onshape. The FeatureScript MCP server lets a model write FeatureScript, insert it into a document, run it, read the errors and revise; it does not edit arbitrary documents. A general Onshape MCP server appears only in the future tense in Onshape's November 2026 post. Onshape's API limits are 2,500 calls per user per year on Free, Standard and EDU Student plans, 5,000 on Professional and 10,000 on Enterprise, pooled at company level; other education plans differ. Onshape's feature page answers "Yes" to whether FeatureScript MCP tool use counts towards that limit.
SOLIDWORKS, Creo and Siemens. Dassault's stated route for third-party agents is the A2A agent protocol, so outside access to desktop SOLIDWORKS goes through the COM API and community servers. PTC's Creo FAQ describes Creo first as an MCP client, then as a server, then headless, without dates. Siemens' Designcenter page says only "Connect your supported AI environment to the Designcenter API through controlled access" beside a screenshot captioned with Claude; no documentation or release number was found.
Community servers fill the gaps, and their quality varies. Star counts show attention, not reliability: freecad-mcp had about 3,400 in November 2026, while one SOLIDWORKS server's README says "most tools have not been validated against a live SolidWorks instance". Beyond the table, there are puran-water/autocad-mcp and U-C4N/Autocad-MCP for AutoCAD; bimwright/rvt-mcp for Revit; catia-v5-mcp-server for CATIA over COM; MCP_CREO_MechDog for Creo via Toolkit; NX_MCP for NX Open; rhinomcp for Rhino; openscad-mcp; tapir-archicad-MCP; and KiCAD-MCP-Server. For Inventor, use Autodesk's own server.
What an agent can actually do through these servers
CAD servers mix typed modelling tools, measurement tools, viewport images and an escape hatch that runs code. What matters more than tool count: can the agent measure its own result, and does the server report failures honestly?
Scroll horizontally to explore all columns →
| Server | Typed modelling tools | Measurement | Viewport images |
|---|---|---|---|
Fusion MCP (official) | Not published; Autodesk lists reading designs, modifying geometry and parameters, exports | Mass, material, geometry extraction | Screenshots from defined views |
Revit MCP (official) | 8 read, 11 write, 39 experimental (move, rotate, sheets, schedules) | Element data, room boundaries | View export |
Zoo MCP (official) | 45 tools: KCL execute, lint, sessions | calculate_mass, calculate_bounding_box_cad, entity_distance | snapshot |
RhinoAI (official) | Commands plus Python and C#; Grasshopper graph tools | Through code | Viewport screenshots |
freecad-mcp (community) | Through code | get_view | |
faust-machines Fusion (community) | About 90: sketches, features, joints, CAM | measure_distance, get_physical_properties, check_interference | render_view |
build123d-mcp (community) | 27 documented, about 50 registered in the source, around one execute tool | measure, validate, find_holes, diff_step | render_view |
What works. Autodesk's list for the Claude connector is reporting, screenshots, bulk edits, data extraction and "parametric parts from a spec", "best with well-defined geometry: extrusions, holes, fillets, shells, and patterns". A MakeUseOf writer using Claude with freecad-mcp got editable brackets, pegboard hooks and enclosures, with bolt patterns and symmetric cut-outs handled well. On a Hacker News thread about LLM-written OpenSCAD, a TPU bike grommet came out close to right on the first print. All are hobby parts with every dimension stated.
What breaks. The same MakeUseOf account describes multi-body assemblies trapping the agent in loops of running code and taking screenshots, which burns tokens. A recurring failure is spatial: the wrong sketch plane, the wrong axis, a fold in the wrong direction. A blogger working out hinge spacing for a wheelchair tray with Claude and CadQuery watched the model fold the tray across an axis that did not exist, then in the wrong direction, before coaching got it to a 16 mm pin spacing that printed and folded flat.
Servers fail in their own ways. In freecad-mcp issue #110, an earlier version of create_object returned success: True for an invalid Pad, and "The model then pads, fillets and exports a shape that does not exist"; the maintainer fixed it in August 2026. In issue #183, since closed, a lost HTTP response made one modifying call run twice. Some limits sit in the CAD API itself: a Solid Edge server notes that assembly pattern and mirror calls are blocked for automation, and an Onshape server documents that section views are unavailable through the REST API.
Why measurement matters more than pictures. The BenchCAD maintainers tested a harness that never showed the model its own renders. It scored the same, −0.002 ± 0.05, at about half the cost, according to their post on 2D proxies (6 October 2026). The reason: in both harnesses the model rendered its part in code, computed how well its silhouettes overlapped the target views and pushed that number up. Most parts came out right, but once the outlines matched, the model could stop with the inside wrong: a duct nearly solid instead of hollow, a wall too thick, a bore straight instead of narrowing.
This is one benchmark team's result with one frontier model, but it matches the freecad-mcp documentation's own warning: "Script success does not certify geometry validity." A useful server therefore lets the agent query a bounding box, a volume or an interference check, and your acceptance test uses those numbers rather than a screenshot.
Security: code execution, local sockets and where your geometry goes
Code execution is a designed feature. The MCP specification says "Tools represent arbitrary code execution and must be treated with appropriate caution" and that the protocol "cannot enforce these security principles at the protocol level". Most CAD servers can run model-written Python, C#, Ruby, AutoLISP or VBA in the CAD process, from AutoLISP in an AutoCAD LT server to Ruby in SketchUp. Autodesk's Inventor tools page states that its Python screening "is not a sandbox", and McNeel's privacy note says its connector "exposes tools that can execute arbitrary code or commands in the Rhino process".
The consequences are concrete. freecad-mcp issue #192, opened on 19 November 2026, describes an agent using execute_code to start Blender on the macOS host through subprocess.Popen with no separate approval. CVE-2026-10688 records a code-injection flaw in the execute_blender_code tool of blender-mcp, a Blender server with about 35,000 GitHub stars. And Invariant Labs showed a poisoned tool description leading Cursor to read SSH keys and pass them out.
Local sockets often have no authentication. Fusion's server has none by design, and the faust-machines Fusion server says directly that "the TCP socket has no authentication". The MCP project's security guidance recommends stdio or an authorisation token for local HTTP servers and names DNS rebinding as a risk for servers left running on localhost. Whether the Fusion endpoint is exposed to that attack has not been publicly tested. Nor has prompt injection through part names, BOM fields or STEP headers: any text a server returns reaches the model, but no CAD-specific demonstration was found.
Before approving a server, fill in the fields of the server table above for it. Practical controls follow: a sandbox seat or virtual machine with no vault access; an allow-list of servers and tools; outbound network limits; logging of every tool call; and read-only tools for the first weeks. Some servers already help: the spkane FreeCAD server hides code tools by default, U-C4N's AutoCAD server has a READ_ONLY mode with allow and deny lists, Revit separates Read, Write and Experimental add-ons, and Archicad keeps "all model-changing actions… subject to explicit user approval". Autodesk's own advice is "Start read-only" and "Save before any modification script".
Where your geometry goes. With an external model, two parties' terms apply. Onshape puts it plainly in a November 2026 post: "The AI provider processes everything the AI model receives: the engineer's prompts and every result an MCP server returns", and "Onshape can't speak for a provider's retention or training policies."
Scroll horizontally to explore all columns →
| Setup | Where the geometry is processed | Training on your data by default |
|---|---|---|
Fusion MCP + Claude on a commercial plan | Your machine, then Anthropic | No for commercial products, per Anthropic; Autodesk says access is governed by Fusion settings and data terms |
Any server + OpenAI API | Your machine or vendor cloud, then OpenAI | No unless you opt in, per OpenAI |
Any server + Gemini API, unpaid tier | Then Google | Yes: content used to "provide, improve, and develop" products, with human review (terms) |
SOLIDWORKS LEO | 3DEXPERIENCE cloud on OUTSCALE | No: models "trained only on proprietary or synthetic data, and not on client data" (vendor claim) |
Creo AI Assistant | PTC-managed Microsoft Azure; no self-hosted LLMs (PTC) | Not stated on the pages cited here |
Local code-CAD + local open-weight model | Your machine | Nothing leaves unless you send it |
On data location, Dassault names OUTSCALE in its July 2026 release but gives no region; Autodesk opened its Forma Building Layout Explorer only to users with data stored in the US. If GDPR or data residency matters, ask each vendor for the region and a data processing agreement before the pilot.
Export control and licences, as the vendors state them. Autodesk's Acceptable Use Policy forbids uploading "export controlled information" and bars replicating its offerings "through training of any machine learning or artificial intelligence algorithm or model". Onshape's Terms of Use exclude "ITAR or EAR controlled" data from commercial accounts, and its AI Advisor is "available for all plans except Onshape Government". Sending controlled geometry to a commercial LLM raises a second export-control question about the LLM provider. No clause specific to MCP or AI agents was found in the Autodesk terms reviewed.
How well do models really do?
Agents that write CAD code and check it in a loop now score high on some benchmarks of fully dimensioned parts, and middling on unseen tasks, assemblies and inferred dimensions. The highest with-tools scores were reported by the vendors of the models being scored, not by the benchmark maintainers.
Scroll horizontally to explore all columns →
| Benchmark (date) | Task | Best reported result | Reported by |
|---|---|---|---|
BenchCAD 1.0, without tools | Four views → CadQuery | 0.747, Claude Sonnet 5.5 (vendor-reported, 1,000-file subset); Gemini 3.1 Pro 0.289 in the maintainers' June 2026 run | Model vendors; maintainers |
BenchCAD 1.0 with tools (Sept 2026) | Same task, 1,000-file subset | 0.959–0.963: GPT-6 Astra, Claude Opus 5.5, Claude Sonnet 5.5 | Model vendors, republished by BenchCAD |
BenchCAD 2.0 (12 Oct 2026) | Six tasks incl. assemblies and ECAD, 100 held-out cases | 0.59, GPT-6 Astra, max effort, about $11/case | Maintainers |
BenchCAD 2.1, customised-harness board (Nov 2026) | Same 100 cases, each vendor's own agent harness | 0.835, Claude Fable 5.2 in Claude Code; 0.825, GPT-6.2 Astra in Codex | Model vendors (Anthropic, OpenAI) |
CADBench (MIT) (May 2026) | One greyscale image → CadQuery | IoU 0.412, valid shapes 0.798, Claude Opus 4.7 | Paper authors |
Text2CAD-Bench (May 2026) | Text → CadQuery, level 3 (complex topology) | IoU 0.23–0.26 for the best general models, with 68–91% invalid outputs (GPT-5.2 had the lowest invalid rate among those four) | Paper authors |
MUSE (v3, Oct 2026) | Specification → multi-part assembly | 57.55% joint success, gpt-6-astra; next best 34.91% | Paper authors |
P3D-Bench (v4, Oct 2026) | Assemblies and parts as programs | Part-matching F1 about 0.52 | Paper authors |
Model names are as the sources print them, and rows are not comparable. BenchCAD's 0.59 is the last 2.0 result run by the maintainers; the November 2.1 scores of about 0.825–0.835 are self-reported by the model vendors, and the maintainers state that "Scores from 2.1 are not comparable with 2.0". On BenchCAD 1.0, Anthropic changed the view resolution mid-series and re-ran earlier models. In the BenchCAD paper, a fine-tuned 2-billion-parameter specialist model reached about 0.75 IoU; IoU is a different measure from the leaderboard score, so the two should not be compared directly.
Three findings change how you would set up a test.
Effort costs money and time. On BenchCAD 2.0, the top two runs cost about $11 per case with a median of 62 minutes (GPT-6 Astra) and about $28 per case at 3.7 hours (Claude Opus 5.5); at low effort the same models scored 0.14 and 0.06. On BenchCAD 1.0, Claude Sonnet 5.5 ranged from about 0.33 at under half a cent per task to 0.963 at about $6.
Assemblies are the open problem. In BenchCAD 2.0's assembly tasks, most zero scores for GPT-6 Astra, Claude Opus 5.5 and Claude Sonnet 5.5 came from correctly shaped parts in the wrong place. MUSE's authors write that "high execution success does not necessarily translate into correct assembly structure or interface geometry". Autodesk Research's neuralCAD-Edit found the best foundation model it tested, GPT 5.2, scored "53% lower (absolute) than CAD experts" on editing tasks in human acceptance trials.
Models read code better than pictures. On the same BenchCAD 1.0 questions, the best score was 0.838 when models were given the CadQuery program and 0.587 when given renders (both Gemini 3.1 Pro variants); a blank image scored 0.375. This tests reading, not tool loops, but it points the same way as the 2D-proxy result.
No benchmark reviewed scores tolerances or GD&T, sketch constraints, feature-tree quality or manufacturability. CADBench's authors state that their test does not require recovery of the "feature tree, sketch constraints, construction order, parameterization, or design intent". "Editable" in these papers means the program runs and produces a valid solid.
A tool builder who launched a SOLIDWORKS add-in on Hacker News in January 2026 wrote that "the LLMs aren't great at CAD right now" and, in a reply, estimated that "probably 75%" of the SOLIDWORKS API scripts (C#, VBA) that models wrote failed to run, a vendor's own estimate with no published method. In a May 2026 thread, users reported failures when the model had to guess dimensions and with snap-fit tolerances, and one user called Fusion's built-in agent "almost comically bad" on four-operation parts. No documented case was found of agent-authored CAD in a certified, safety-relevant or volume-production part.
In SimScale's 2026 survey (published March 2026) of 350 senior engineering leaders at companies with more than 1,000 employees, autonomous agents were in use in 11% of design and CAD workflows; SimScale sells simulation software. IoT Analytics, with 120 respondents, found 87% expect AI built into CAD and PLM, while no more than 3% "fully trust" AI output for any activity. Autodesk's vendor-commissioned AI Pulse says 98% of leaders use at least one AI tool, and its 2027 State of Design & Make says 40% have integrated AI into design workflows.
A first experiment: three paths
The aim is one bounded test on parts your team already makes, judged by measurement, with assemblies and released designs out of scope.
Path A: a team pilot (recommended plan)
- Pick the route by package. Fusion: Fusion MCP, read-only first. Onshape: FeatureScript MCP or a REST community server, with the API quota in view. SOLIDWORKS or Inventor: the built-in assistant, or a server on a non-production seat. No CAD seat, or STEP is the deliverable: build123d or CadQuery.
- Baseline. Take 10 to 20 fully dimensioned parts modelled in the last quarter (brackets, plates, enclosures, turned parts) and record current modelling and review time per part.
- Inputs. A written spec per part with every dimension stated, mating geometry as STEP where relevant, and your modelling conventions.
- Success criterion, set in advance and checked by script. Bounding box and key dimensions within tolerance; volume or mass within a set percentage; hole count and positions; a kernel validity check; no interference with mating parts; review time below the baseline.
- Failures to log. Wrong plane or flipped feature; features reported as created that are invalid; operations repeated after a timeout; dropped threads and chamfers; token use climbing in screenshot loops; Fusion Compute's session limits.
- Cost, fully counted. Model API spend per part including failed runs; Flex tokens if the Fusion Assistant is used; Onshape API calls; supervising engineer time.
- Stop rule. A fixed number of rounds per part, then a person takes over. Assemblies stay manual in round one. "Wait and re-test in six months" is a valid result, and so is keeping the current workflow.
Path B: security sign-off before the pilot
Settle six things before the first session: a copy of the design, kept apart from the vault; an AI provider on a commercial or API plan your organisation has approved, which is Onshape's own advice; no export-controlled data; read-only tools first; a saved model before any write; and code-execution tools only behind an approval step where the server supports one. Record the server-table fields for each server you plan to allow, and ask the CAD vendor and the model provider for their data region and processing terms.
Path C: solo or maker
For 3D-printed or hobby parts, the cheapest start is FreeCAD with a community server, or a code-CAD library with no CAD seat. OpenSCAD is easy for models to write, but it outputs meshes only, and its last stable release is from 2021 (a 2026.11 release candidate appeared on 7 November 2026). CadQuery (v2.8.0, June 2026) and build123d (v0.13.0, September 2026) produce exact B-rep geometry, export STEP, run headless and are Apache-2.0 licensed; move to them as soon as a part will be machined, drawn or edited in another CAD. Ask for several views plus a measurement after each step, and state every dimension.
A selection of AI components by route
Each component below works with a named CAD package, has a documented status and was checked against a primary page. Vendor assistants and MCP servers are in the tables above. This is a selection, not a ranking, and most capability statements are the vendor's own.
Built into the CAD
Scroll horizontally to explore all columns →
| Component | Package | What it does | Status (date) |
|---|---|---|---|
Fusion | Rebuilds a feature timeline from history-less geometry | Tech Preview, Design Extension (Oct 2026) | |
Fusion | Sheets, views, auto-dimensioning strategies | GA | |
SOLIDWORKS | Drawings, assembly structure, BREP-to-parametric, VBA macros | Beta labels in 2026 releases; reseller calls it Early Access | |
Solid Edge 2026 | 2D drawings "70–80% complete" (vendor) | Available (Oct 2025) | |
NX X Manufacturing | A generative machining suggestion per feature | Release 2606 (June 2026) | |
Onshape | ML transfer of drawing annotations | Beta |
SOLIDWORKS' BREP-to-parametric skill works on single bodies and, per the reseller, "does not recover the original feature tree or design intent". Fusion's text-to-geometry has not shipped: Autodesk's 2026 roadmap says it is "working toward" it.
Single-job plug-ins and AI-native tools
Scroll horizontally to explore all columns →
| Component | Works with | Job | Status (date) |
|---|---|---|---|
Fusion, Mastercam, NX, GibbsCAM, SolidCAM | 3-axis and 3+2 CAM programming | Available; vendor claims 1,000+ shops (Aug 2026) | |
Fusion | DFM, estimating, AI toolpaths | CAM last described as beta (Jan 2025) | |
SOLIDWORKS, Inventor | Drawings "80–90% complete" after training on a company's drawings (vendor); no GD&T | Available | |
Fusion | Sketch constraint and dimension suggestions | Free, v1.1.0 (Dec 2025) | |
Fusion add-in, web | Scan or mesh → feature tree | GA (3 Aug 2026) | |
Files and PLM; push from Creo and NX | Drawing review (AI), 3D checks (deterministic) | AutoReview 2.0 (July 2026) | |
SOLIDWORKS, Creo, NX, CATIA, Inventor | DFM, tolerance stack-up, standards checks | Sold to enterprises | |
Any CAD, via PDF or image drawings | Reads drawings into JSON; API and remote MCP | Available, public pricing | |
SOLIDWORKS | Agent: text, image or STEP → parts, drawings | v2 (7 Oct 2026) | |
Own CAD; STEP export | Zookeeper agent writes KCL that builds B-rep | Zookeeper from 5 Feb 2026 | |
Onshape, Fusion; desktop bridge to others | Copilot; hosted MCP and REST API | MCP and API in beta |
Independent tests are rare. The engineering.com hands-on of Backflip's GA release was mixed: an adaptor ring failed four out of four times in Fast mode but looked flawless in Thinking mode, and no knob reconstruction was fully accurate. Make:'s look at Hestus, marked as sponsored, found its suggestions strong on a symmetric enclosure and out of its depth on a clock face. A Xometry test of seven text-to-CAD tools from August 2025 (free tiers, three dimensioned prompts; republished by Wevolver in June 2026) found Zoo produced an accurate, editable tube but an inaccurate gear and no manifold block. The test predates Zoo's Zookeeper agent.
Not CAD: mesh generators. Tools such as Meshy output FBX, OBJ, GLB or STL with no STEP, so no parameters, feature history or exact faces.
Beside CAD: physics AI. Simulation surrogates such as Ansys SimAI take tessellated geometry ("VTP (recommended) or STL") and solver results; CAD models reach them only as file exports.
AEC and BIM
Autodesk's Revit server launched in June as a read-only Tech Preview with seven tools; it now has Read, Write and Experimental add-ons with 8, 11 and 39 tools, none of which executes code, according to Autodesk's tool reference; the experimental set can write without per-action confirmation, as AEC Magazine reports. Revit now labels AI changes in Undo and logs Assistant runs in its journal. DynamoMCP lets the Assistant place, wire and run Dynamo nodes; outside clients are not documented.
Scroll horizontally to explore all columns →
| Tool | Official MCP (status) | What agents can do |
|---|---|---|
Revit | Technical Preview, Revit 2027.2 | Query, select, move, rotate, sheets, schedules, IFC export |
Beta; Collaborate or Max only | Ten tool groups; every change needs approval and is one undo step | |
Technology Preview, 2026.2+ | Bentley lists it with STAAD.Pro as the first in a planned rollout | |
MCP add-on | Prompt site plans; TestFit's solver makes the geometry | |
SketchUp | None; community server | Ruby code execution |
A BIMsmith blog tried six scenarios with Claude and Revit, hit context limits on a 15-room analysis and got one stair-code finding wrong; its method is not disclosed, and commenters dispute which server it used.
Electronics
No incumbent board-design vendor has shipped an official MCP server for its layout tool. Altium lists servers for Altium 365 and Altium Designer as "Coming Soon". Siemens says its Fuse EDA AI system "supports Model Context Protocol", with capabilities "available in forthcoming releases".
Scroll horizontally to explore all columns →
| Tool | Agent access | Note |
|---|---|---|
Official MCP (Oct 2026), OAuth | Reads components, nets, BOM, layout; hands jobs to the Flux agent | |
Community, about 250 tools | Schematic, placement, routing, DRC | |
"Custom API & MCP" (enterprise) | Learned routing on a placed board | |
File upload, no MCP | Placement and routing candidates |
In a KiCad forum thread, an agentic server routed a three-component board with no DRC errors, and replies said it would be faster by hand. A JLCPCB walkthrough built a 5 A buck converter with Claude Code: KiCad's ERC only warned when two overlapping net labels shorted the current-sense and current-limit nets, and a pin-by-pin netlist check caught it. The author notes that "Zero DRC errors mean the layout meets its rules. It does not confirm that the schematic matches your intent." BenchCAD 2.0's ECAD task found 85–96% of components and 68–80% of connections, and its authors add that "Neither run ever tested for a short or an open".
FAQ
Can ChatGPT or Claude do CAD?
Yes, through an MCP server, the vendor's API or a code-CAD library such as CadQuery: dependably for queries, reports, bulk edits and fully dimensioned simple parts, weakly for assemblies. Autodesk calls its Claude–Fusion connector "not yet reliable for generating complex CAD models or full assemblies".
Is there AI in AutoCAD?
Yes. Autodesk Assistant in AutoCAD 2027 is a tech preview that checks drawings against your organisation's CAD standards file, according to Autodesk's AutoCAD 2027 post; Autodesk says this first release is focused on standards. Autodesk's AutoCAD MCP server runs inside that Assistant; outside agents use community servers.
Is there an MCP server for SOLIDWORKS?
Not from Dassault, which points third-party agents to the A2A protocol. Community servers drive SOLIDWORKS over the COM API on Windows; check each one's validation status.
What is the difference between an MCP server and the CAD API?
The MCP server wraps the API so an LLM client can discover and call it; it cannot do anything the API cannot. Autodesk says its servers execute only what the AI application requests.
Can AI generate editable CAD, or only meshes? Can I turn OpenSCAD output into STEP?
Code-CAD libraries such as CadQuery produce exact B-rep geometry and export STEP. OpenSCAD produces meshes only, so a true STEP file means rewriting the part in a B-rep tool.
What AI can make 2D drawings from a 3D model?
Fusion's Drawing Automation is generally available; LEO drawing generation in SOLIDWORKS carried a Beta label in 2026; Solid Edge, Onshape and DraftAid have their own. None handles GD&T fully; DraftAid says tolerances are added by hand.
Is my design data used for training? Is it safe for proprietary designs?
With an external model, both the CAD vendor's and the AI provider's terms apply, and Onshape notes that it "can't speak for a provider's retention or training policies". Use an approved commercial or API plan and keep export-controlled data out.
Will AI replace CAD designers?
The evidence does not point that way yet: in SimScale's 2026 survey, autonomous agents were used in 11% of design and CAD workflows, and no public benchmark scores tolerances, design intent or manufacturability.
Is there free AI CAD software?
Hestus is a free Fusion add-in; FreeCAD, its community MCP servers and CadQuery are open source; Fusion's personal plan includes 10 Flex tokens a month for the Assistant. An external agent's model still costs money unless run locally.
What is the difference between generative design and generative AI in CAD?
Generative design is usually topology optimisation, a solver that removes material against loads; Autodesk's page for the Fusion Simulation Extension, which carries generative studies, does not mention AI. Generative AI uses trained models such as LLM assistants.


