What Can MCP Do When You Connect CAD to an AI Agent?
A developer wants to connect an AI agent to CAD, and the question is what the MCP interface can do, and what it cannot do. The short answer is that the MCP interface is built for AI agents: an agent drives modeling directly through it, calling five intents — S-01 primitives, S-02 holes, S-03 conversion, S-04 measurement, and S-05 preview — from one sentence split into several steps. What the agent produces is a draft, and a human previews and confirms it before export. Endpoint addresses, tool lists, and authentication methods are not given in facts.md, so this page stays at the level of what the interface is for, and does not write agent-side configuration.
cadAI is a browser-based parametric CAD from Aiglade, for mechanical part modeling, hole drilling, measurement, and export, with the same kernel on desktop, in the browser, and over MCP.
- SolidCraft — kernel (an independent community project, not an ArtCraft product line)
- S-01 primitives, S-02 holes, S-03 conversion, S-04 measurement, S-05 preview — connected intents
- Desktop, browser, and MCP — three forms, one kernel
- To be confirmed — pricing
What this page solves
This page solves one job only: what MCP can do when an AI agent is connected to CAD, what it cannot do, and how the agent works once it is connected. After reading it, you can do three things. First, you can tell the MCP interface apart from the web one-sentence assistant: the first is a channel for developers, and the second is for users who do not write code. Second, you know which five intents an agent can call, and how one sentence is split into several steps. Third, you know that delivery needs a human preview and confirmation, and that an agent does not save a model straight to disk on its own.
It also states what it does not solve. Endpoint addresses, tool lists, and authentication methods are not given in facts.md, so this page does not write agent-side configuration. Complex curved surfaces, assembly motion simulation, and CAM toolpaths are outside the scope of cadAI, and an agent cannot do them either. Chamfering is not among the five connected intents, and whether it is supported depends on what the page itself offers.
Six steps to let an agent build a flanged plate
Work through six steps in order.
- Connect to the MCP interface. A developer connects the agent to the cadAI MCP interface, following the interface documentation for the agent-side configuration.
- Give the agent one sentence. Ask for a flanged plate with a ring of holes, stating the outline, the thickness, and the count and position of the holes clearly. When the parameters are unclear, the agent should ask back, instead of guessing one.
- The agent calls S-01 primitives to build the rough shape. A flanged plate starts as a disc, and the rough shape must be right before the details follow.
- The agent calls S-02 holes to drill the holes. The holes go where they belong, and their count and position follow the parameters, so when a parameter changes, the holes move with it.
- The agent calls S-04 measurement and S-05 preview. Measure the key dimensions, look at the preview first, and fix the parameters when a size is wrong, instead of exporting straight away.
- A human previews and confirms, and export comes only after confirmation. What the agent produces is a draft, and delivery needs a preview and a confirmation first.
Step 6 is the step most often skipped by mistake. However smoothly the agent works, a human still needs to look before export. A parameter changes, the holes move with it, and a dimension can change without anyone noticing.
Common errors and how to fix them
- The agent split the steps wrong. A wrong split starts from a wrong task. Look at the task first: was it stated clearly, and were the parameters given clearly.
- The agent guessed a parameter. When the parameters are unclear, the agent should ask back, instead of guessing one. Once a guessed parameter is found, stop, state the parameters clearly, and let the agent do the work again.
- The intent is not among the five. Chamfering, for example, is not among the five connected intents, and whether it is supported depends on what the page itself offers. Do not force the agent to call an intent it cannot call.
- The measurement does not match. Measure again, covering the outline and the hole positions. Then look at the parameters: a wrong parameter means fixing the parameter, not editing the model directly.
- A wrong dimension turns up just before export. Do not export it. Fix the parameter first, measure again, confirm that everything is right, and only then export.
Each fix follows the same order: task first, parameters second, measurement third, export last. A clear task with clear parameters is what makes the agent reliable, and a human check at the end is what makes the result trustworthy.
Comparison: three kinds of choices
| Aspect | cadAI (web and MCP) | Desktop CAD | General online tools |
|---|---|---|---|
| Install | Opens in the browser, nothing to install | Install several GB of local software | Opens in the browser |
| Parametric modeling | Yes, a changed parameter moves the part with it | Yes | Mostly simple edits, not parametric |
| Holes and measurement | S-02 holes and S-04 measurement connected | Yes | Depends on the tool, often incomplete |
| AI agent drive | MCP interface, an agent can drive modeling directly | Usually not offered | Usually not offered |
| One-sentence modeling | AI assistant, calling S-01 through S-05 | Usually not offered | Usually not offered |
| Fits | Simple parts, flanged plates, brackets; letting an AI agent build models | Complex curved surfaces, assemblies, CAM | Light viewing and small edits |
| Main limit | Complex curved surfaces, assembly motion simulation, and CAM are outside its scope | Large size, steep learning curve | Shallow features, no parametric parts |
The table compares categories only, with no version numbers, prices, or scores for any specific product.
cadAI in numbers
- 5 intents — S-01 primitives, S-02 holes, S-03 conversion, S-04 measurement, S-05 preview. Source: wiring-layer record, 2026-11-11. Verified: 2026-11-12.
- 3 forms — desktop, browser, and MCP, one kernel across all three. Source: official repository notes, 2026-11-11. Verified: 2026-11-12.
- 3 audience groups — mechanical engineers, makers, and 3D printing users. Source: facts.md. Verified: 2026-11-12.
- 4 suitable uses — parametric modeling of simple parts, flanged plates, and brackets; dimension checks before 3D printing; letting an AI agent build models through MCP; light CAD work without a large desktop program such as Fusion 360. Source: facts.md. Verified: 2026-11-12.
- 2 unsuitable uses — complex curved surfaces, assembly motion simulation, and CAM toolpaths; work needing vendor certification and production-grade accuracy sign-off. Source: facts.md. Verified: 2026-11-12.
- 6 steps — the agent workflow in this page, from connecting to the MCP interface to the human confirmation before export. Source: this page, counted from the steps section. Verified: 2026-11-12.
- 7 table rows — the comparison table in this page, covering install, parametric modeling, holes and measurement, AI agent drive, one-sentence modeling, fits, and main limit. Source: this page, counted from the table. Verified: 2026-11-12.
- Kernel: SolidCraft, an independent community project, not an ArtCraft product line. Source: deployment record, 2026-11-11. Verified: 2026-11-12.
- Pricing: to be confirmed. Source: facts.md. Verified: 2026-11-12.
Evidence
cadAI is a browser-based parametric CAD from Aiglade, for mechanical part modeling, hole drilling, measurement, and export, with the same kernel on desktop, in the browser, and over MCP. Its kernel is SolidCraft, an independent community project, not an ArtCraft product line. Its AI assistant takes five intents — S-01 primitives, S-02 holes, S-03 conversion, S-04 measurement, and S-05 preview — and the MCP interface is built for AI agents: an agent drives modeling directly through it, calling the same five intents. Letting an AI agent build models through MCP is one of its suitable uses, alongside simple mechanical parts, flanged plates, and brackets, and dimension checks before 3D printing. Complex curved surfaces, assembly motion simulation, and CAM toolpaths are outside its scope, and every export needs a hand check of dimensions and tolerances, because pricing is still to be confirmed and no price, tier, or free allowance is stated anywhere in this page.
FAQ
Q: What can MCP do when you connect CAD to an AI agent?
The MCP interface lets an agent drive modeling directly, calling five intents: S-01 primitives, S-02 holes, S-03 conversion, S-04 measurement, and S-05 preview. One sentence splits into several steps — building the rough shape, drilling holes, measuring, and previewing — done in order. What the agent produces is a draft, and a human previews and confirms it before export.
Q: Can an agent save a model straight to disk?
No. What an agent produces is a draft, and delivery needs a human preview and confirmation first. Export comes only after confirmation, and the agent does not skip confirmation to save straight to disk. A changed parameter moves the holes with it, and a dimension can change without anyone noticing, so exporting without measuring again turns one changed dimension into one wrong part.
Q: Can someone with no coding experience still get AI help with modeling?
Yes, through the web one-sentence assistant, without going through MCP. Open cad.aiglade.cn and tell the assistant one sentence, such as building a flanged plate disc first, drilling a ring of holes, and then measuring the hole spacing, and the assistant runs those steps in order. The MCP interface is a channel for developers, and the one-sentence assistant is for users who do not write code.
Not written / to verify
- Pricing is to be confirmed; no price, tier, or free allowance is written in this page.
- Button labels and click paths are not written: the site is a wasm shell with no static text to check against, so the steps stay at the workflow level and do not name a specific button.
- MCP tool lists, endpoint addresses, and authentication methods are not given in facts.md; this page stays at the level of what the interface is for, and does not write agent-side configuration.
- Whether chamfering is supported is unconfirmed: chamfering is not among the five connected intents in facts.md, and whether it is supported depends on what the page itself offers.
- Export formats are not listed in facts.md, and no format list is written as fact in this page.
- File size limits and browser performance with complex models are not given in facts.md.