agentcad
Open-source MCP server that lets a coding agent write Python CAD scripts, render previews, and export STEP, STL, GLB, or OBJ parts.
If your agent keeps writing geometry that won't slice, agentcad is a cheap fix: free, local, Apache-2.0, and built around feeding validity metrics back to the model. Pre-execution validation, structured JSON metrics, spec-check against a written checklist, and diff mode with an overlap score cover the four places agent CAD usually breaks. The tradeoff is that it is a backend, not a modeling app. Pick it when the agent is the designer and you are the reviewer.
Verified 4d ago · liveness 71/100 · cite: rightaichoice.com/tools/agentcad
- AI agents iterating on 3D part geometry with an automated validity feedback loop
- Developers automating parametric CAD generation with LLMs and Python
- Engineers who want printable STL and STEP output straight from a coding agent session
- Hobbyists exploring prompt-to-part workflows locally without an account
- Designers who want interactive GUI modeling with a mouse and viewport
- Teams needing real-time collaborative editing or shared cloud workspaces
- Anyone without basic Python literacy — the scripts are the interface
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip agentcad if you want to model interactively in a viewport with a mouse — this is an agent-facing CLI and MCP server where Python scripts are the interface, with no GUI editor.
The default install is build123d only; adding CadQuery pulls a separate dependency tree with its own licensing surface to review
agentcad is free and open source under Apache-2.0 with source on GitHub, so cost is zero and the only outlay is your own machine time. That places it in a different budget line from subscription GUI suites like Fusion 360 or Onshape, and from commercial prompt-to-CAD services that bill per generation. The maintenance cost is real though: one maintainer, no SLA, and CadQuery support shipped as an optional extra rather than baked in.
In short
agentcad — Open-source MCP server that lets a coding agent write Python CAD scripts, render previews, and export STEP, STL, GLB, or OBJ parts. Best for AI agents iterating on 3D part geometry with an automated validity feedback loop, Developers automating parametric CAD generation with LLMs and Python, Engineers who want printable STL and STEP output straight from a coding agent session. Free to use.
What's new in agentcad
Checked 4 days agoAcross the latest 4 updates: 2 feature updates and 2 changelog entries.
Using Meta's Muse agent for 3D printing and CAD
Meta's Muse agent found agentcad on skills.sh and produced a 3D-printable iPhone Duo bedside dock in about four minutes, exporting STL and USDZ.
Measuring agentcad's effectiveness
CADGenBench comparison shows how agentcad changed GPT-5 Nano's design quality, validity, and cost.
Introducing diff
Compare CAD versions with side-by-side renders, a colored overlay, and measurements that help an agent verify an edit.
Introducing parts
Give assembly pieces names, colors, and groups so people and agents can inspect how they fit together.
Viability Score
How well maintained and how widely used is agentcad? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →Key Features
- MCP server exposing CAD design, render, and export tools to coding agents
- Works with Claude Code, Cursor, WindSurf, and other MCP-capable agents
- CLI for local script execution and versioned STEP file output
- build123d-native with CadQuery compatibility as an optional extra
- Pre-execution static validation under 100 ms (syntax, missing show_object, unresolvable imports)
- Structured JSON metrics per run on stdout: dimensions, volume, surface_area, face_count, edge_count, is_valid
- Four-view PNG preview composite (top, bottom, upper iso, lower iso) per run
- Export to STEP, STL, GLB, and OBJ for 3D printing and web viewers
- Diff mode compares two versions with side-by-side renders, colored overlay, and overlap score
- Parts mode names, colors, and groups assembly pieces so agents and humans can inspect fit
- Spec-check mode measures finished geometry against a written JSON checklist
- Topology report via `agentcad inspect` for debugging broken geometry
- Zero-boilerplate scripts with pre-injected build123d primitives, show_object, and edit helpers
- agentcad init installs a Claude Code skill plus an AGENTS.md (or CLAUDE.md) workflow block
- Runs locally with no signup — files never leave your machine
About agentcad
agentcad is an open-source MCP server and CLI that hands CAD work to your coding agent instead of your mouse. You install it with `pip install agentcad[mcp]` on Python 3.10 to 3.12, run `agentcad init --name myproject`, and paste a prompt into Claude Code, Cursor, Windsurf, or any MCP-capable agent. The agent writes a Python script using build123d (CadQuery as an optional extra), calls `show_object()` on a shape, and agentcad runs it locally, producing a versioned STEP file, geometric metrics, a four-view preview PNG, and a review viewer HTML. It is aimed at developers, mechanical engineers, and hobbyists who want prompt-to-part output they can actually print, without an account or cloud upload. The feedback loop is the product. Every run returns structured JSON on stdout with `dimensions`, `volume`, `surface_area`, `face_count`, `edge_count`, and an `is_valid` flag, while progress lines go to stderr so a parser doesn't choke. Static validation catches syntax errors, a missing `show_object()`, and unresolvable imports in under 100 ms without consuming a version number. If the geometry is wrong, `agentcad measure` reads feature sizes back off the finished file and `agentcad inspect` gives a topology report. Scripts need no imports, since build123d primitives, `show_object`, and agentcad's edit helpers are pre-injected, so a part can start at `box = Box(10, 20, 5)`. Export covers STEP, STL, GLB, and OBJ, including USDZ for iPhone delivery in the Meta Muse walkthrough. Named parts (July 2026) let you color and group assembly pieces; spec-check mode (June 2026) measures finished geometry against a written JSON checklist; diff mode (August 2026) compares versions with side-by-side renders, a colored overlay, and an overlap score. Meta's Muse agent reportedly went from finding agentcad on skills.sh to a printable iPhone Duo bedside dock in about four minutes. Source is on GitHub under Apache-2.0 (github.com/jdilla1277/agentcad), built by James Dillard, and it runs
Behind the Verdict
The pitch worth testing is the loop, not the tool count. Static validation runs in under 100 ms and, more importantly, does not burn a version number, so an agent can fail cheaply and retry. Metrics then come back as JSON the model can reason about: volume, surface area, face and edge counts, an `is_valid` flag. In practice, that is what stops a session from spiraling into twenty near-identical broken parts. Since June 2026, spec-and-measure has been the piece I'd reach for first on anything with real tolerances. You write the dimensions you wanted as a checklist; agentcad measures the finished file and tells the agent where it drifted. Diff mode (August 2026) pairs side-by-side renders with a colored overlay and overlap score, which is the fastest way to confirm an edit only changed the feature you asked about. Parts mode (July 2026) adds names, colors, and groups so multi-piece assemblies stay legible to both humans and agents. Pick this when the agent is doing the geometry and you are reviewing the result, when you want STEP and STL out of the same script, or when files cannot leave a local machine. Pass when you want to drag a fillet by hand, need real-time collaborative editing, or expect a vendor SLA. It is also a poor fit if nobody on the team reads Python, because the script is the interface. Closest alternative is wiring build123d or CadQuery into your own agent harness. That is viable, and you would own validation, versioning, previews, and export. agentcad's value is that those four are already wired together and it costs nothing. Two caveats. The September 2026 CADGenBench figures (48% higher GPT-5 Nano score, 2.5x valid outputs, 21.8% lower run cost) are the vendor measuring itself, so reproduce them on your own parts before quoting them. And the
Researching agentcad? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Real-world workflow fit
Concrete scenarios for the personas agentcad actually fits — and what changes day-one when you adopt it.
Paste a prompt into Claude Code asking for a phone stand 80mm wide, 50mm deep, with a 5mm lip at 60 degrees; the agent writes build123d Python, runs `agentcad run script.py --label first`, and reads back `metrics.dimensions` plus `is_valid` from the JSON.
Outcome: A versioned STEP file in `v1_first/` alongside a four-view preview PNG and a viewer HTML, ready to slice — no manual CAD session needed.
Ask the agent for a 60x40x25mm rounded enclosure with a snap-fit lid and 2mm walls, then ask it to edit one wall; from v2 the run also emits `diff_side.png` with a colored overlay and overlap score.
Outcome: The agent verifies its own edit against the previous version and re-runs only if the diff shows unintended geometry change, so revision history stays auditable.
Work through the Wright Flyer case study — 234 parts written in Python against the Christman measured drawings — using `agentcad parts` to name, color, and group subassemblies.
Outcome: A multi-part assembly where each piece is inspectable in the review viewer and any part can be version-diffed independently.
Use Cases
- Design a phone stand with specific dimensions and export it for 3D printing
- Iterate on an electronics enclosure with snap-fit lid by comparing version diffs
- Reconstruct a historical object like the 1903 Wright Flyer (234 parts) from Christman measured drawings via prompts
- Generate a vase profile with variable thickness and validate volume constraints
- Automate CAD design for multiple part variants via script parameter overrides
- Inspect a STEP file for geometric errors before sending to fabrication
- Build an assembly of named, colored, grouped parts so an agent can check fit
Models Under the Hood
as of 2026-10-17
Limitations
- agentcad supports Python 3.10 to 3.12 only; the OpenCascade bindings it depends on do not exist for 3.13 yet.
- Installation is via pip or uv, and the default install is build123d only, with CadQuery as an optional extra that carries its own dependency and licensing surface.
- Structured JSON responses go to stdout while progress lines go to stderr, so anything that merges the two streams will corrupt the parse.
- It also requires an MCP-capable coding agent or manual CLI use to drive the CAD workflow.
as of 2026-10-04
Verification history
We have re-verified agentcad 10 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
- — re-verified summary, description, our verdict, our analysis, pricing model, pricing tiers, features, integrations, who it suits, who should skip it
Showing the 6 most recent of 10 verification passes.
Free to cite with attribution — this page re-verifies continuously.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Plans compared
For each published agentcad tier: who it actually fits, and what it adds vs. the previous tier. Cross-reference the cost calculator above for projected annual outlay.
Open Source
$0
Ideal for
Developers, hobbyists, and engineering teams who want agent-driven CAD without a subscription, running entirely on their own machine
What this tier adds
Starting (and only) tier — Apache-2.0 licensed, free, no signup, all functionality including MCP server, CLI, validation, diff, parts, and spec-check
Where the pricing makes sense
The company stage and team size where agentcad's pricing actually pencils out — and where peers do it cheaper.
agentcad is free and open source under Apache-2.0 with source on GitHub, so cost is zero and the only outlay is your own machine time. That places it in a different budget line from subscription GUI suites like Fusion 360 or Onshape, and from commercial prompt-to-CAD services that bill per generation. The maintenance cost is real though: one maintainer, no SLA, and CadQuery support shipped as an optional extra rather than baked in.
Setup time & first value
How long it actually takes to get something useful out of agentcad — broken out by persona, not the marketing-page minute.
Developer with Python 3.12 already installed: about five minutes from empty folder to printable part — create a venv, `pip install 'agentcad[mcp]'`, `agentcad init --name myproject`, paste the phone-stand prompt into Claude Code. Engineers with an existing CadQuery or build123d codebase: longer, because you'll want to walk the `agentcad docs` output contract before trusting an agent to read
Switching to or from agentcad
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From CadQuery scripts: install with the CadQuery extra and reuse existing geometry code, since agentcad is CadQuery-compatible
- →From build123d scripts: drop the boilerplate imports — primitives and show_object are pre-injected, so a part can start directly at `box = Box(10, 20, 5)`
- →From manual Fusion 360 iteration: keep Fusion for interactive work, then hand the final dimensions to an agent to regenerate as parametric Python under version control
- →From OpenSCAD: rewrite geometry in build123d Python; agentcad's metrics feedback replaces the render-and-eyeball loop
- ↗To Fusion 360 or Onshape: export the versioned STEP file from `output.step` and import it for interactive editing or cloud collaboration
- ↗To another slicer or printer host: export STL, GLB, or OBJ directly from the run outputs for 3D printing and web viewers
- ↗To a custom pipeline: consume the structured JSON on stdout — dimensions, volume, face_count, is_valid — and skip the CLI wrapper entirely
Integrations
Resources & Guides
Tutorials & Learning
YouTube returned 6 videos for “agentcad”, and we withheld 6: 6 could not be judged, because “agentcad” is a single word that other videos use for other things. We are showing none, because we could not prove any of them are about agentcad.
Official links
Tools that pair well with agentcad
Common stack mates teams adopt alongside agentcad, with the specific reason each pairing earns its keep.
Text To Cad
Open-source agent skills that turn text or image prompts into CAD models, STEP files, G-code, and robot descriptions.
VibeUE
Open-source UE 5.8 plugin that hands AI agents direct control inside the Unreal Editor
Foyr Neo
Browser-based interior design software that turns a floor plan into photorealistic 4K renders and 3D walkthroughs in three steps
Featured Head-to-Head Comparisons
Agentcad vs Polycam
Polycam is the best choice for professionals needing photorealistic 3D scans and floor plans from real-world spaces. Agentcad is ideal for AI agents and developers who want to programmatically generate and iterate on 3D parts. Choose Polycam for reality capture; choose agentcad for agent-driven CAD design.
Agentcad vs Cognition Ai
Choose Cognition AI's Devin if you're an enterprise team needing autonomous software engineering with measurable productivity gains and a financial guarantee. Choose agentcad if you're a developer or AI agent building 3D parts programmatically with open-source, scriptable CAD tools. They solve fundamentally different problems: software engineering vs. mechanical design.
Agentcad vs Bito
Bito and agentcad serve completely different domains. Bito excels as a system-wide context layer for AI coding agents in multi-repo software projects, providing architectural awareness and automated scoping. agentcad is a focused, free open-source tool for AI-driven 3D part design using Python scripting. Choose Bito if you need cross-repo AI assistance; choose agentcad if your agents need to generate CAD files.
Alternatives to agentcad
View all
Text To Cad
Open-source agent skills that turn text or image prompts into CAD models, STEP files, G-code, and robot descriptions.
Frequently Asked Questions
Used agentcad? Help shape our editorial sentiment research.