Using hackerscad with an AI agent

hackerscad is new, so it is not in the training data of the models: a model left to guess writes OpenSCAD or CadQuery instead. Give it both:

• the skill: what hackerscad is, how to work with it, the rules that avoid most mistakes, and which tools the MCP server offers, so the agent knows they are there and what to use them for;
• the MCP server: the tools themselves, with which the agent checks models, looks at renders, measures, exports and drives the open application.

The skill alone works too, through the command line; the MCP server alone leaves the agent to discover the language from the tools.

The easy way: ask your agent

Write this to your agent: it installs the skill, then follows the skill's instructions to install hackerscad from the releases if it is missing (it asks you first) and to add the MCP server.

The skill keeps itself up to date: it carries the stamp of the docs it was made from, checks it against hcad docs --stamp and, when hackerscad has been updated, regenerates itself with the installed hcad.

By hand

1. Find hcad

Both run hcad, the command line tool, and the agent starts it by itself from any folder. hcad --version tells whether it is in your PATH. If it is not, either put it there or use its full path in the commands below:

2. Add the skill

The rules that avoid most mistakes (SKILL.md, or AGENTS.md for tools that want a rules file) and the whole reference (llms-full.txt; llms.txt is a short overview). hcad prints them, so they always match the installed version; in Claude Code, for example:

mkdir -p ~/.claude/skills/hackerscad
hcad docs --skill > ~/.claude/skills/hackerscad/SKILL.md
hcad docs > ~/.claude/skills/hackerscad/llms-full.txt

The folder for each agent:

3. Add the MCP server, too

One command, for all your projects; then restart the agent. In the application, the main menu's AI Agents gives these commands with the path of your installation, shows which agent drives the window and what it asked, and what to ask your agent to check it is connected (it then shows there by its own name): an agent working on the window that never shows there is not using hackerscad's tools (it may be clicking through the screen, as Codex's computer use does on macOS). Name it hcad, as below: the skill tells the agent its tools (hcad_check, hcad_render…), and one server is enough (two, under different names, give the agent every tool twice).

Without hcad in the PATH, the full path takes the place of the second hcad, the one before mcp. For example, with the AppImage in ~/Apps:

claude mcp add -s user hcad -- ~/Apps/hackerscad-nightly-x86_64.AppImage hcad mcp
codex mcp add hcad -- ~/Apps/hackerscad-nightly-x86_64.AppImage hcad mcp
agy mcp add hcad ~/Apps/hackerscad-nightly-x86_64.AppImage hcad mcp

(the name of the AppImage changes with every version: the link of step 1 does not), or on Windows:

claude mcp add -s user hcad -- C:\path\to\hackerscad\bin\hcad.exe mcp

Other tools take the same server in their MCP settings: the command (hcad, or its full path) with the argument mcp.

Then ask, for example: "a hexagonal pen holder, 10 cm tall, with a rounded bottom, as 3MF".