wheresryan22/anatomy

This page introduces an external open-source repository. It is not an HDATF product.

What it does

Anatomy is an Agent Skill (a folder with a SKILL.md file that an AI coding agent reads as instructions) written for Claude Code and also loaded by OpenCode, Codex and Antigravity. It has the agent explain a technical idea by inventing a physical machine whose working is the idea and drawing it as an interactive isometric figure, and it is meant for people who need explanatory figures for docs, blog posts or landing pages. The input is a request that names the idea or the object with its real numbers, such as sizes, counts and rates, together with the output format and a light or dark theme. The output is source the user owns, either a Node build script with a standalone HTML page, a single SVG file, or React and Next.js geometry modules for React 18 or later, with the drawing kit included. There are three kinds of figure: SVG made of solid parts in a card with a live readout, SVG plus a WebGL shader layer for fire, water or steam, and a 3D mode in beta whose moving parts are re-projected every frame. SKILL.md has the agent first pick a size (a Figure of 40 to 80 parts, a Hero of 80 to 150 parts, or an Epic of 6 hours or more built only when the user chooses it), then collect the true numbers and build the geometry with the kit. Verification is part of that order, with kit/audit.mjs checking whether any part passes through another and whether every end sits on a mount, and the skill telling the agent to read every PNG taken with headless Chrome. It uses only built-in modules of Node.js 20.10 or later, so there are no API keys and nothing to npm install, the checks need Chrome or Chromium, and a Codex sandbox that blocks Chrome leaves only the checks failing. It is installed with the skills CLI or by cloning into the agent's skills folder under the MIT license, and installing only makes the skill available through /anatomy, while each figure still needs a long agent session. The README limits the subjects to machines, instruments, tools and furniture and leaves data charts with axes and series, people and exact maps outside its scope. According to SECURITY.md, the check scripts start headless Chrome with a throwaway profile and the allow-file-access-from-files option, and drive.mjs runs JavaScript passed to it inside the page. On request, publish.mjs sends a figure's source and prompt to the public anatomy-community repository as a pull request after the user agrees, which needs a signed-in GitHub CLI, and it reads time and tokens from the Claude Code session log.

License

MIT Permissive. Commercial use and changes are allowed if the copyright notice is kept.

More in this category

• storytold/figcraft
  FigCraft is an interface design app written in Rust by the ArtCraft team (the storytold organization on GitHub), built as a clean-room reimplementation of the Figma workflow for designers and developers who make app and web screens. It covers auto layout frames, components and variants, variables with modes, and prototypes, while its Dev Mode writes code such as CSS, React and SwiftUI in which component properties become parameters. Files are .figcraft archives on the local disk, layers export to PNG, SVG, PDF and other formats, .fig files open on a best-effort basis, and .fig export is marked experimental. Every menu item and tool runs a command from one list of more than 400 commands, and the same commands are reached from the figcraft-cli tool, an MCP server (a protocol that lets AI agents call tools) and a JSON control channel. The MCP server runs headless or attaches to a running app, and the control channel accepts connections only on 127.0.0.1, but no key check was found in the control code read for this research. The app works offline with no account or telemetry, but according to the README the Online option in the Share menu uploads the file by default to a public server (collab.craftapps.dev). That server stores the file unencrypted and readable by the operator until it is removed or nobody opens it for 30 days, with the edit or view link itself serving as access. A self-hosted figcraft-server (a preview, with Linux builds in the releases) keeps a team's files behind edit and view links and has no user accounts yet. Getting started means using a release package for macOS, Windows, Linux or FreeBSD, the static web build or a build with Rust 1.95 or later, and connecting an agent is a separate step that adds figcraft-cli mcp to the MCP client's settings. The Korean interface catalog is a machine draft that awaits review by native speakers, Korean text and IME input work in documents, and the web build needs a browser with WebGPU or WebGL2 and offers Korean only when built with embedded CJK fonts. The project calls itself alpha, puts itself about half way to replacing Figma for real work (49 percent in the README, 53 percent in the ROADMAP), lists checking against real Figma files as open work, and states that four to six AI agents working in parallel built it in about 25 hours. The GitHub license field shows Apache-2.0, while the LICENSE files and README offer MIT or Apache-2.0 at the user's choice, with the ArtCraft name and logos excluded from that license.
• Jakubantalik/transitions.dev
  transitions.dev is a repository that combines a collection of CSS transitions for web interfaces with a CLI that adds them to projects, a skill for coding agents, and transitions-agent, which checks interface motion. Its main users are front-end developers who want consistent screen transitions and animations, and people who build interfaces with coding agents such as Claude Code or Cursor. Each transition is provided as a copy-ready CSS snippet with :root custom properties, t- prefixed classes and prefers-reduced-motion handling. The transitions-dev CLI (MIT, Node 18 or later) adds and lists free transitions without an account, while Pro transitions are downloaded from api.transitions.dev after a device sign-in, with entitlement checked on every download. transitions-agent (MIT) scans interface code locally with 11 rules and gives a motion score from 0 to 100, and its GitHub Action posts a score comment on a PR and then opens a fix PR against that PR branch. The fix command sends the affected files to a hosted fix service, shows the changes Claude generates as a diff first, writes them to files only after the user approves, and is described as limited to CSS transitions and animation guards. The free sign-up tier offers unlimited scans and Action runs, unlimited fixes through the user's own Claude Code or Cursor, and up to 10 hosted fixes a month and 2 a day, while the Business plan at 59 dollars a month includes 5 Pro seats and 200 fixes a month. The custom Transitions.dev License allows modifying and shipping in personal and commercial projects and forbids redistribution as a competing library, and company use of Pro transitions or paid Agent features needs a Business seat for each person. The CLI and agent tooling are MIT, and GitHub's NOASSERTION license label appears to come from this custom license. Transition counts differ across sources, with 12 in the README table, 18 in the README text, 43 or more in the repository description, and 32 free plus 11 Pro in the agent document, so the actual number should be counted directly, and because hosted fixes send code to an outside service, the scope of files to send should be set before using it on company code.
• CopilotKit/OpenIntelligentUI
  OpenIntelligentUI is a chat application and development foundation that answers with text, tables, and interactive interfaces. It targets developers offering explanations, comparisons, and small tools within conversations. A Python agent processes the request and streams results through CopilotKit and AG-UI to the browser. Jev selects the presentation, using A2UI for basic tables and sandboxed HTML, CSS, and JavaScript for richer visuals and calculators. Controls can send selected values into a follow-up request, but the documentation provides no API for directly patching a previous output from another call. Requirements include Node.js 22+, pnpm 9+, Python 3.12+, uv, and separate OpenAI and Jev keys. Visitor keys remain in browser memory but pass through the application server to providers, so the deployment operator must be trusted. The inspected query_data source returns the entire sample CSV without query filtering or live company data access. Conversation checkpoints live in process memory and are not durable across restarts or shared across multiple servers.
• storytold/vectorcraft
  VectorCraft is a Rust vector illustration editor for designers working with paths, text, layers and multiple artboards. It opens its JSON-based native document format, SVG and PDF, including PDF-compatible Illustrator files, and exports vector or raster artwork. Geometry operations, live appearances and a separate rendering engine let users edit shapes without making the interface itself the document model. The CLI and MCP, a tool connection protocol for agents, expose drawing commands, document inspection and rendered previews in either headless or connected desktop mode. The current Cargo manifest specifies Rust 1.95, while desktop targets include macOS, Windows, Linux and FreeBSD and the browser build uses WebAssembly. Japanese fonts can come from the optional shared font repository, with installed system fonts used when that input is absent on desktop. The README still identifies gaps in advanced effects, CJK vertical composition, interaction fidelity and Windows/Linux packaging rather than promising full Illustrator parity. The MCP documentation separately warns that progress notifications, in-flight cancellation and resource subscriptions are unimplemented, and rendering speed claims have not been independently reproduced here.
• kargulstudio/sales-crm
  The README describes a Next.js 16, React 19 and Tailwind CSS 4 starter. Includes scripts for AVIF image conversion and poster generation from video and Rive files.

Only repositories in the ranked Trendshift lists are included, and the lists are used only to find candidates. We do not copy their ranks. Descriptions, licenses and star counts come from each GitHub repository. The notes are our own reading. We have not tested these projects, and a place on a trending list does not prove quality.