presentation-king

Agents draft. You finish.

A BYO-agent presentation tool for developers who make slides with coding agents (Claude Code, Codex, …). Your agent writes the deck over MCP; you finish it by hand in a local web editor — drag, snap, restyle — without typing another prompt. Free, open source, no built-in AI: the reasoning always comes from the agent subscription you already pay for.

Status: pre-release — not on npm yet. Everything below works today when run from source; see INSTALL.md. Once published, the npx presentation-king … commands become the real interface.

Why

• Stop describing pixel nudges in prose ("move that box left… no, the other box").
• Human edits and agent edits coexist without destroying each other.
• One renderer: what you edit is exactly what exports to HTML and PDF.

Quick start

Full guide for running from source, MCP registration, and the Claude Code skill/hook: INSTALL.md.

# 1. Create a deck (pick one of three design sets)
npx presentation-king init my-talk

# 2. Let your agent draft it (MCP, see below), or edit right away
npx presentation-king serve my-talk

# 3. Ship it
npx presentation-king export my-talk        # static bundle + PDF

How it works

• deck.json is the source of truth — a JSON scene graph: fixed 1280×720 canvas, absolute coordinates, stable element ids, typed elements (text / image / shape / html-embed), design-token styles. HTML is only ever a render artifact.
• One deck server per deck (.pk/server.json) owns the state; every write — your edits, agent ops, undo — goes through a single serialized write queue.
• MCP is the agent's primary path. The stdio MCP server is a thin client that auto-starts the deck server headless, so apply_ops works even when no editor is open. Direct file edits are an exception path: the schema validator quarantines broken elements, the editor renders them visibly broken, and saving never destroys them.
• Blast radius: pin a selection in the editor and the agent's apply_ops is rejected at write time if it touches anything outside the pin. No pin = whole deck writable.
• Two undo lanes: Cmd+Z only walks your edits. Agent changes are checkpoints, reverted only via checkpoint restore.
• Change digest: the agent reads a semantic summary of what you changed since its last checkpoint (get_user_changes / ack_user_changes), so it never bulldozes your manual polish.
• Diff view: inspect every agent checkpoint and each of your edit sessions as a semantic diff — change list, canvas highlights, and a before-state preview toggle.

Agent setup (MCP)

Register the MCP server with your agent, per deck:

# Claude Code
claude mcp add presentation-king -- npx presentation-king mcp /path/to/my-talk

Tools: get_deck, get_selection, apply_ops, get_screenshot, get_guidelines, get_user_changes, ack_user_changes, wait_for_prompt, get_deck_info. Agent-facing usage guide: agent/SKILL.md.

Waiting mode

Have the agent open the editor and listen: the agent calls wait_for_prompt; whatever you type into the editor's prompt bar (plus your pinned selection) is delivered straight to it. The tool times out with no_prompt (default 55 s) — the agent simply calls it again to keep waiting.

Session-start hook (optional)

Inject unconsumed user changes at the start of every Claude Code session — add to your project's .claude/settings.json:

{
  "hooks": {
    "SessionStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "node node_modules/presentation-king/hooks/inject-digest.mjs my-talk"
          }
        ]
      }
    ]
  }
}

The script is standalone and silent when there is nothing to report.

Design sets

Three built-in sets — minimal-light, dark-tech, bold-business — each with four layers: design tokens, an element style dictionary, layout templates, and agent instructions. Decks reference tokens (token:accent, style heading-1), so switching themes restyles every slide without moving a single element. Raw hex values remain available as an escape hatch.

User guides (2 layers)

• Global: ~/.presentation-king/guide.md — rules for all your presentations
• Per deck: <deck>/guide.md — audience/tone for this one

Both are editable in the editor's Guide panel and merged into get_guidelines together with the active design set's agent instructions.

Editing features (v1)

Move/resize with snap guides (drag threshold, cursor feedback) · inline text editing · html-embed source editing · multi-select (shift-click, containment marquee) · right-click context menus (canvas + slides) · element clipboard (Cmd+C/X/V) · image insert via file, URL, drag-&-drop, or paste · token-scoped styling with hex escape hatch · z-order · slide add/duplicate/delete/drag-reorder with range/toggle multi-select · zoom (fit/50/75/100%) · human-lane undo/redo · align/distribute · play mode (fullscreen + keyboard) · presenter view (notes + next slide) · agent chat sidebar with checkpoint diff cards, selection pin, and waiting-mode prompts · change diff view.

Non-goals (v1): hosting/sharing service, PPTX, realtime collaboration, animations, freeform drawing, cell-level table editing (use html-embed), desktop app, built-in AI, mid-deck layout-reflow switching.

Development

pnpm install
pnpm dev              # editor (5173) + deck server on the sample deck
pnpm build && pnpm typecheck && pnpm lint
pnpm test:schema && pnpm test:history && pnpm test:mcp
pnpm test:export && pnpm test:hook
pnpm test:e2e         # Playwright

Monorepo: packages/schema (scene graph, validator, ops, history, diff) · packages/design-sets · packages/renderer (single renderer for editor = export = PDF) · packages/server (deck server + MCP) · packages/editor (React) · packages/cli.

MIT.