plantt - a beautiful planning Gantt backbone for large AI projects, LLM-driveable and front-end only

Making sure that compute capacity is appropriately used and that ambitious AI projects aren't blocked by preventable delays isn't easy, and planning tools are woefully hard to manipulate, read, and keep up to date (and ugly). Plantt addresses all of these with a lightweight LLM-driveable (via a skill) Gantt chart generator and editor. Easily share plans with collaborators, who can make their own changes and share them back with just a URL. Front-end only, to make both privacy and sharing natural.

Live: http://openathena.ai/plantt/

Features meant to make coordinating large model development plans easy and automatable

Lightweight, shareable, agent-driveable

• Agent-remote controllable via a local LLM relay, so larger text plans can be automatically translated into JSON plans, and edits can be made easily.
• Direct-manipulation in-UI editing — drag to move/resize, snap, wire dependencies, double-click to edit via modal.
• Embedded JSON editor as the source of truth, with live validation.
• Plans shareable via compressed URLs.
• Undo/redo history tree (branches preserved), named plans.
• Themeable (Tufte, Solarized, LaTeX, Catppuccin, Nord, Gruvbox, Dracula, Rosé Pine, Print, …), following the OS light/dark preference. Themes are local-only and importable/exportable as JSON.

Model planning-focused

• Compute-capacity lanes: per-cluster utilization, FLOPs-scaled lane heights, over-subscription. Activities can rely on clusters, and overutilization is flagged.
• Dependency connection with a violations highlighted.

Develop

npm install
npm run dev      # http://localhost:5173
npm run build    # → dist/  (deployed to GitHub Pages by .github/workflows/deploy.yml)

The app is plain ES modules bundled by Vite for hot reload; src/main.js is the whole app.

Remote control (optional)

Loading the app with ?agent=1 lets a local tool drive the active plan over a small localhost relay — see .claude/skills/plantt-remote. It exposes a window.plantt API (describe, getState, outline, getDeps, getDependents, setModel, an atomic name-addressed apply(ops), and a themes namespace); a normal visit (without ?agent=1) exposes nothing reachable.

Installing the skill

The skill is one self-contained folder — .claude/skills/plantt-remote/ — holding SKILL.md (the instructions) and relay.mjs (the localhost bridge). It has no build step and no runtime dependency on this repo: it drives a running plantt tab and pulls the live schema from the app itself (GET /schema). It needs only Node.js (to run the relay) and curl.

Claude Code / any agent that reads ~/.claude/skills/:

• Working inside this repo — nothing to do. Project skills under .claude/skills/ are auto-discovered whenever the working directory is in the plantt checkout.
• Globally (every project) — symlink it into your user skills dir so it stays in lockstep with the repo copy:

  ln -s "$PWD/.claude/skills/plantt-remote" ~/.claude/skills/plantt-remote

  Use cp -R instead for a detached copy — but a copy won't track schema/op changes, and this skill's value is staying in sync with src/schema.js, so prefer the symlink.

Any other shell-capable LLM agent (no ~/.claude/skills convention): SKILL.md is plain Markdown — load or paste its contents into the agent's context as instructions. That's the entire interface; the agent then runs node relay.mjs and drives it with curl exactly as the file describes. (Portable toolkits can also vendor the folder and copy it into ~/.claude/skills/ on setup.)