agentcad

agentcad

CAD tool for AI agents. Give your coding agent the ability to design 3D models.

Your agent writes build123d Python scripts by default. agentcad handles execution, STEP export, PNG rendering, mesh export (STL/GLB/OBJ), geometric metrics, validation, diffing, and browser preview. CadQuery remains available as an explicit compatibility mode. Each command’s final response is structured JSON on stdout.

Reading the output: the JSON response is written to stdout; human-readable progress and diagnostics go to stderr. Parse stdout as JSON and treat stderr as plain text — don’t merge the streams with 2>&1 before a JSON parser, or the progress lines will break parsing. If you need both, capture them separately.

agentcad is open source under the Apache License 2.0. It runs locally and requires no signup.

Demo

A coding agent designing in agentcad, live. See more at agentcad.dev.

On CADGenBench: 48% higher CAD Score and 21.8% lower model cost

In a GPT-5 Nano comparison, the agent using agentcad also produced 2.5 times as many valid outputs. Watch the video, read the benchmark results, or view the change log for more recent updates.

The v0.7.0 release brings the five-stage workflow below to the installed guide and built-in help. In its Candidate L CADGenBench run, GPT-5 Nano scored .245 overall (.175 generation, .353 editing) with 80 of 81 valid outputs. That is a single run, not a new matched direct-build123d comparison; see the experiment closeout for the baseline, caveats, and next evaluations.

Quick start

Install agentcad, then paste this into Claude Code, Cursor, or any coding agent:

Create a Python 3.12 virtual environment, then:

pip install agentcad
agentcad init --name phone-stand

Design me a phone stand: a simple angled cradle that holds a phone
at 60 degrees. About 80mm wide, 50mm deep, with a 5mm lip at the bottom
to keep the phone from sliding. Show me a preview when you're done.

To install the reusable agent skill separately, use npx skills add jdilla1277/agentcad-skill or clawhub install agentcad. Both install guidance, not the CAD executable; install agentcad with pip as shown above. For a project-local guide, run agentcad skill install instead.

Keep generated files in a build directory

Create agentcad.toml in your source project before initialization:

build_dir = "./build"

Then run agentcad init. Version directories, the generated manifest, viewers, exports, and logs go under build/; authored scripts and installed guidance stay in the source project. Add the build directory to your own .gitignore.

For CI, initialize and select an independent root with agentcad init --build-dir /tmp/cad-build and agentcad run model.py --label first --build-dir /tmp/cad-build. Overrides do not edit project configuration. Relative build paths resolve from the project root, not the caller’s working directory. Existing projects keep their current layout unless configured. --output remains a deprecated version label, never a destination. See agentcad docs artifacts for the full contract.

Required agent workflow

agentcad init installs a guide that requires the same workflow for new models and edits:

Inspect → Specify → Build → Validate → Deliver

• Inspect: understand drawings, requirements, orientation, dimensions, and any source geometry before writing a model.
• Specify: record measurable requirements and assumptions. For edits, also identify the target, expected change, and geometry that must remain stable.
• Build: create the simplest geometry that can satisfy the specification; use the smallest bounded operation for an edit.
• Validate: inspect visual, dimensional, difference, and structural evidence. If a check fails, return to the earliest incorrect stage and iterate.
• Deliver: hand off only the exact STEP and requested exports that passed.

Capabilities by workflow stage

Inspect

• agentcad inspect output.step — bounded topology, orientation, addressable IDs, edit risk, and structural evidence
• agentcad measure output.step — dimensions, edge lengths, face areas, and circular/cylindrical feature measurements
• agentcad render output.step --view top,front,right,iso — visual evidence for orientation and feature location
• agentcad parts list REF / parts show REF ID — inspect named parts and isolate one versioned part

Specify

• spec.json — records supported cylindrical-feature and structure requirements before modeling
• agentcad check-spec output.step spec.json — evaluates that contract during validation

Build

• agentcad run script.py --label label — execute a build123d script, producing a versioned STEP file and geometric metrics
• Pre-injected helpers — build primitives, import existing STEP geometry, select features, and perform bounded or validated editing operations

Validate

• agentcad run ... --preview — four-view PNG for visual verification; the browser viewer can export an on-demand turntable GIF
• agentcad diff 1 2 — compare versions, including actual shared/reference-only/candidate-only source-frame volume for valid closed solids
• agentcad view old.step new.step — open a synchronized A/B comparison with separate centered projection and source-frame 3D volume artifacts
• Live project viewer — successful runs open one stable local project URL and refresh it after subsequent completed builds. Leave the tab open while iterating; camera and compatible review settings survive updates. Failed builds leave the last successful model visible. Versioned viewer.html snapshots remain available; from v2, previous and current are preloaded for side-by-side, overlay, and Parts-tab change review (--no-view opts out)
• Spatial review comments — pin human feedback to a surface and named part; comments persist locally for the agent’s next revision
• agentcad review list --status open — read shared review threads; humans and agents can reply, resolve, and reopen with an auditable history
• agentcad parts view REF — hand off an isolated, focused, or grouped part review viewer

Deliver

• agentcad run ... --export stl,glb or agentcad export — write mesh deliverables and report validation for each written format
• agentcad viewer [open|status|stop] — open the live project or manage its lightweight local service; open is the default
• agentcad docs [section] — runtime-aware built-in documentation and worked examples

Spatial review comments

Live project viewers opened after a run or with agentcad viewer open use the existing service bound to 127.0.0.1. Click Comments in the bottom-left viewer controls, then + New comment (or press C). Click the model, confirm the inferred named part and revision scope, then choose Send comment. To hold feedback back, choose Save draft instead; unsent drafts can be edited, deleted, sent individually, or submitted together with Send all drafts.

Comments are plain local JSON under the selected build root’s .agentcad/reviews—there is no database or hosted account. On its next turn, an agent can discover and reply to the review:

agentcad context
agentcad review list --status open
agentcad review comment --message "Could this rib be thinner?" --part support_rib --version current
agentcad review reply C1 --message "Increased the clearance to 4 mm." --version current
agentcad review resolve C1 --message "Implemented in the current revision." --version current

Agents can also initiate an immediately-open thread on a named part. Add --scope previous|both to target comparison sides or --point-mm X,Y,Z to place its pin more precisely. Humans and agents can both reply, resolve, or reopen. Every action records its actor, timestamp, optional message, and associated revision. The older mark-addressed command remains available for compatibility.

New builds wait while you write or send a comment or reply, then update the model once you send, save, or cancel. Current/Previous/Both follow the displayed revision sides; the original source version stays in the thread. If a part disappears, its comment remains in the list as unplaced. Hiding a part hides its comment pins in the 3D view; its threads stay in the sidebar. Showing the part restores its pins. Comments become available to the agent on its next turn; they do not start a new agent run. File viewers and temporary part/diff viewers remain read-only.

To reposition the model, click Pan and drag, or hold Shift while dragging. Click Pan again to return to orbiting. Fit view centers the visible model and adjusts the zoom to fit it on screen.

--label names the version; the JSON response returns the actual file under outputs.step. The older --output LABEL spelling remains a deprecated compatibility alias and never denotes a destination path.

Keep one preview open

After a normal run or import, give the human project_viewer.url from the JSON response. That URL tracks the latest completed viewer; viewer still points to this version’s snapshot. An active page prevents another browser tab opening. Updates are checked about once a second (less often in background tabs). A new model replaces the previous one only after it loads successfully.

--no-view never launches a browser or starts the viewer service. If it still generates viewer artifacts, an already-open project page can receive them. The core-only --no-preview --no-diff --no-view combination generates no viewer; the page retains the older model and reports that the new preview is unavailable.

The service binds to 127.0.0.1 and serves only registered viewer snapshots at private project URLs. These are local bookmarks, not public share links. URLs survive service restarts and builds, but moving the project folder changes its identity. Viewer state lives in ~/.cache/agentcad/viewer (override with AGENTCAD_VIEWER_HOME for isolated automation). Keep that directory to preserve the saved port and private URLs. If the port is occupied, AgentCAD reports the problem without silently changing the URL. Use agentcad viewer status and agentcad viewer stop for diagnostics; agentcad viewer open restarts it. An incompatible service protocol requires stopping the service with the installation that started it before reopening with the new installation.

No boilerplate

Scripts need zero imports. By default, build123d primitives, show_object, and agentcad edit helpers are pre-injected:

box = Box(10, 20, 5)
show_object(box)

For editor completion or explicit dependencies, the same AgentCAD-owned callables are available from the stable authoring module:

from agentcad.api import load_step, safe_cut, translate, show_object

Import build123d primitives such as Box or Vector from build123d itself. Do not import STEP writers: show_object hands the result back to AgentCAD, which writes and tracks the canonical STEP file.

In build123d scripts, capture each shape object once, including across show_object(), show_assembly(), and show_compound() calls and within assembly iterables. Capturing an assembly also captures all its descendants; do not capture a child separately or reuse it in another captured assembly. Repeated references fail with error_kind: "duplicate_capture"; changing the name or ID does not create another instance. For intentional instances, use from copy import deepcopy, then copy the shape before positioning and capturing it. See agentcad docs parts for an example.

agentcad init records build123d as the project runtime. That keeps the script API, built-in docs, and subsequent runs on one clear default.

CadQuery compatibility

CadQuery remains supported for existing scripts and projects, but it is not the default authoring path and it is not installed by default. pip install agentcad ships build123d and the OpenCascade binding only. Enable the compatibility runtime with the optional extra:

pip install "agentcad[cadquery]"
agentcad daemon restart   # if a daemon was already running

The extra pulls in CadQuery and its own dependency chain (CasADi and the COIN-OR solver libraries it bundles), which carries a larger third-party license surface than the default install. Selecting the CadQuery runtime without the extra returns a structured error naming this command. Adopting agentcad inside a product or hosted service? See Incorporating agentcad into your project.

For a CadQuery project:

agentcad init --name legacy-model --runtime cadquery
agentcad docs quickstart --runtime cadquery
agentcad run script.py --label first

For a one-off CadQuery script inside a build123d project:

agentcad docs preamble --runtime cadquery
agentcad run legacy.py --label legacy --runtime cadquery

Keep each script on one CAD API. Conflicting imports, including mixed imports, are rejected against the project runtime before daemon startup or version creation, with guidance for fixing the script or intentionally overriding the runtime. Unpinned projects still detect the runtime and reject ambiguous scripts. Run responses identify the selection in runtime_source (command, project, or detection, which includes the default fallback). Run agentcad docs runtimes for the complete dispatch contract.

MCP integration

For native tool integration with Claude Code, Cursor, or Windsurf:

pip install agentcad[mcp]

Extras combine: pip install "agentcad[mcp,cadquery]" installs both.

Add to .mcp.json:

{"agentcad": {"command": "python", "args": ["-m", "agentcad.mcp"]}}

For local STEP/STP/BREP files, call import_cad with file, cwd, and initialize=True to start a project. The default response includes the import JSON and up to nine inspection PNG attachments. Use report_images=False for an image-free report or report=False to keep the bundle on disk. Follow the returned IDs and paths with inspect and run. Local paths refer to the MCP execution filesystem. See agentcad docs mcp for controls.

For retry-safe imports, pass idempotency_key="supplier-001" (CLI: agentcad import supplier.step --init --idempotency-key supplier-001). Preserve the key and identical source/options after a lost response: retries return the original version without importing or opening the viewer again. A different source, label, runtime request, validation profile, report/image limits or viewer/diff options conflicts with that key. Explicit checks JSON content and --defer-checks must also match. Use a new key for an intentional new import. Keys are scoped to the selected build directory; the original file must remain readable with identical bytes.

A contended retry waits up to ten seconds, then reports in_progress if the original process still owns execution; retry the same request later. A stopped process releases ownership automatically. Retrying recovers a committed version and marks unfinished artifacts interrupted, or explicitly reports that execution stopped before commit. Optional analysis is not automatically resumed. Saved responses describe the original attempt; use inspect for fresh inspection. When retrying from another project subdirectory, returned actions target the selected baseline from that caller directory. Inline detail/media may be trimmed further to keep the final report within its text budget; saved analysis and IDs stay unchanged. If required references cannot fit, inspection_delivery reports the omission and provides a retrieval action with a larger allowance. Deleting .agentcad/import_requests deletes retry protection for those keys.

Imports now return a durable inspection_job and wait for inspection by default. Use --inspection-background or --inspection-wait-timeout SECONDS to return before analysis finishes. A wait deadline leaves the worker running; follow the returned status/wait/resume/cancel actions. These wait controls do not change import retry identity. See inspection jobs for CLI/MCP examples, checkpoints and cancellation behavior.

Import and inspect --report default to a 2,000-estimated-token limit for the complete JSON response, with bounded image samples and a full-response retrieval action. --full-response opts out of whole-response compaction. Agents can supply --context-window, --response-budget, --response-text-budget, --image-budget and a caller-estimated --vision-profile to bound complete inspection delivery. Limited responses include a full saved response and an executable retrieval action. See inspection budgets for estimates, CLI/MCP parity and the distinction between whole-response limits and report preparation. Saved response snapshots default to 30 days and 64 MiB per build root, with configurable limits and explicit expiry/eviction responses. Cleanup preserves geometry, issued IDs, import retry records and resumable jobs. Analysis, image sheets and completed job results also have shared cache limits, defaulting to 30 days/512 MiB per build root. Resumable checkpoints remain protected; a full quota stops enrichment with a resume action while preserving the imported model.

Cloud projects

For setup and a complete agent build-and-edit loop, see the cloud workflow guide.

If your host supplies a hosted MCP URL, a supported agent can connect directly with the issued key pair, without installing AgentCAD locally. Call cloud_help for remote tools and explicit content uploads. The guide covers URL setup and retrieving private evidence through MCP.

If your host has issued an AgentCAD API address and key pair, the same local MCP server also offers cloud_run, cloud_import, cloud_measure, cloud_inspect, cloud_check_spec, cloud_render, and cloud_diff. Project, job, and download tools support that workflow. Set AGENTCAD_API_URL, AGENTCAD_KEY_ID, and AGENTCAD_KEY_SECRET in the server process environment. Keep credentials in a password manager, outside project files and agent prompts.

For the local cloud CLI, install pip install 'agentcad[cloud]==0.7.2' (also included in agentcad[mcp]), as shown in the guide above:

agentcad cloud projects create --name Bracket
agentcad cloud run PROJECT examples/cloud/bracket.py --label first
agentcad cloud job PROJECT JOB --wait
agentcad cloud measure PROJECT 1
agentcad cloud inspect PROJECT 1 --summary
agentcad cloud check-spec PROJECT 1 examples/cloud/spec.json
agentcad cloud render PROJECT 1
agentcad cloud download PROJECT iso.png --job RENDER_JOB --kind render

Replace PROJECT and JOB with the returned IDs. Submissions return immediately; poll the same job until it completes. Keep its idempotency_key for an identical retry after a lost response. Upload an edited script with a new key and label to create another version, then compare with agentcad cloud diff PROJECT 1 2.

Only the named script, STEP, or spec file is uploaded. No local project is needed. Downloads require an explicit destination and preserve existing files by default. Job completion, geometry validity, spec verdicts, and human approval remain separate. Run agentcad docs cloud for all commands, limits, and MCP tool names. Cloud access uses your host’s installed CAD engine and credentials; it is not a hosted remote MCP endpoint or a general remote mode for local commands.

Requirements

• Python 3.10–3.12 (OpenCascade bindings do not support 3.13+)

License

Apache License 2.0. See LICENSE.

Feedback

If your agent struggles, run agentcad feedback "what happened" to capture a friction log.