NoPPT

Open-source, Gamma-style card presentation platform — generate slide decks with AI and export to HTML / PDF / PNG.

English · 简体中文

NoPPT is an open-source alternative to Gamma / SlidesAI. Describe your topic (or supply reference material) and NoPPT plans a structured deck, then renders self-contained, interactive HTML slides with a card-style design system. It runs as a monorepo: a React + Vite editor and a NestJS backend that orchestrates LLM providers (OpenAI / Anthropic / Ollama / Free.ai / v0 by Vercel).

✨ Features

• AI deck generation — topic → structured plan → interactive HTML slides, with optional reference HTML / image / text as grounding material.

• Card-style design system — consistent, responsive cards, theming, icon sets, and dense / compact layouts.

• Full editor — drag-and-drop canvas, element inspector, slide list, AI chat assistant, live preview.

• Multi-language UI — switch between Simplified Chinese and English from Settings → Interface; the language is persisted to the server config and also drives the language of generated content and server messages.

• Flexible export — single-file HTML, ZIP of static web assets, PDF, and PNG sequences.

• Extensible model routing — configure multiple providers per stage (planning / content / editing), with high-contrast and audit modes.

• MCP server for AI agents — 8 noppt_* tools let any MCP-compatible client (Hermes, Claude Desktop, …) generate, edit, export and preview decks programmatically.

• Private & self-hostable — all data lives under packages/server/data/; no cloud lock-in.

Related MCP server: PowerPoint MCP Server

🧱 Tech Stack

📂 Project Structure

packages/
├── core/      # Shared domain types and utilities (@noppt/core)
├── ai/        # AI agent orchestration (@noppt/ai): planning, HTML slide generation, audit
├── audit/     # Visual / content self-audit engine (@noppt/audit)
├── server/    # NestJS backend (MCP + REST) (@noppt/server)
└── web/       # React editor and presentation viewer (@noppt/web)

📚 Repo Wiki

Developer-facing architecture and per-package API docs (generated by code analysis, committed to git): see repo-wiki/.

🛡️ AI Quality Assurance (Inline Review · LLM / VLM Review · Auto-Regeneration · Auto-Fix)

NoPPT doesn't just generate slides — it verifies them. Two layers of checking wrap the generation pipeline, both configurable under Settings → AI Model → AI Audit Settings and AI Inline Self-Check Settings.

Inline review (per-slide, during generation)

Controlled by inlineSelfCheckSettings. As each slide is produced, NoPPT runs two immediate checks before moving on:

• LLM critique (llmCritique) — reviews the slide copy for quality, typos, and semantic duplication.

• VLM placeholder check (vlmPlaceholder) — renders the slide and uses a vision model (VLM) to catch failed AI-image placeholders / obvious layout breakage.

When a slide's issue count exceeds threshold (default 7), it is auto-regenerated in place, up to maxRetries (default 1) times — no user action required.

Post-generation multi-engine audit (LLM + VLM review)

Controlled by auditSettings. After the full deck is assembled, the @noppt/audit engine runs four sub-engines — layout, visual, content, fidelity (plus sanitization) — driven by:

• LLM review (llmReview) — an LLM grades content quality / correctness.

• VLM review (vlmReview) — a vision model screenshots each slide and grades the visual design / layout.

The combined score is gated against thresholds.pass / thresholds.warn. Strictness presets (strict / normal / relaxed) tune the pass thresholds to 80 / 70 / 50.

Auto-regeneration

When the audit still finds blocking issues that can't be auto-fixed, NoPPT auto-regenerates the affected slide(s), feeding the LLM / VLM feedback back into the generator, up to maxRegenerationRetries (default 1) times, then keeps the best-scoring version.

Automatic post-processing fix

Controlled by autoFix (default true). Layout / style issues flagged as fixable are patched directly by the AutoFixer (deterministic HTML / CSS transforms — e.g. fixing icon spans, overflow, contrast) and then re-verified. Issues that can't be auto-fixed fall through to the auto-regeneration path above.

Configuration

// packages/server/data/config.json
"auditSettings": {
  "enabled": true, "strictness": "normal",
  "autoFix": true,              // automatic post-processing fix
  "maxRegenerationRetries": 1,  // auto-regeneration ceiling
  "llmReview": true,            // LLM review
  "vlmReview": true,            // VLM review
  "engines": { "layout": true, "visual": true, "content": true, "fidelity": true, "sanitization": true }
},
"inlineSelfCheckSettings": {
  "enabled": true, "llmCritique": true, "vlmPlaceholder": true,
  "maxRetries": 1, "threshold": 7       // per-slide auto-regeneration
}

Requires an audit (text) and auditVlm (vision) model routed under Settings → AI Model; if a model is missing, that review stage is skipped with a warning rather than failing the run.

🚀 Quick Start

Requirements: Node ≥ 18.17.0, pnpm ≥ 8.0.0.

# 1. Install dependencies
pnpm install

# 2. Start both frontend (5173) and backend (3001) in dev
pnpm dev:all

# Or run them separately
pnpm dev          # frontend only (http://localhost:5173)
pnpm dev:server   # backend only (http://localhost:3001)

Open http://localhost:5173. Configure a model provider under Settings → AI Model before generating.

Build & Production

pnpm build         # build all packages
pnpm build:web     # build frontend only
pnpm build:server  # build backend only
pnpm start:server  # run the built backend (PORT=3001 by default)

Scripts

🌐 Internationalization (i18n)

• UI language is controlled in Settings → Interface → Interface language (zh-CN / en). It is persisted to data/config.json and also sent via the Accept-Language header, so server-side error messages follow the same setting.

• Generation language can be set independently per generation in the AI generate dialog (defaults to following the interface language). It is forwarded to the AI agents so the generated slide copy is produced in the chosen language.

• The frontend uses a lightweight, dependency-free dictionary (packages/web/src/i18n/); the backend keeps a parallel message catalog (packages/server/src/i18n/).

🔌 MCP Server & AI Agent Integration

NoPPT ships a built-in MCP (Model Context Protocol) server, so MCP-compatible agents (Hermes, Claude Desktop, or your own client) can drive the whole deck lifecycle — generate → poll → edit → export → preview — purely through tool calls.

Tools (8)

Results come back as JSON in result.content[0].text. Business errors set result.isError = true while HTTP stays 200; only auth failures return HTTP 401.

Connect a client (Hermes example)

# .env — keep the secret out of version-controlled config
MCP_NOPPT_API_KEY=nppt_<32-hex>

# config.yaml
mcp_servers:
  noppt:
    url: http://localhost:3001/api/mcp
    enabled: true
    headers:
      Authorization: Bearer ${MCP_NOPPT_API_KEY}
      # optional: override the user scope
      # X-User-Id: local

Then run hermes mcp test noppt — it should report Tools discovered: 8. MCP configuration is only loaded when a session starts, so restart the session after changing it.

Issue an API key

Option A — fixed dev key in packages/server/data/server.env (auto-provisioned on first boot):

NOPPT_DEV_KEY=nppt_<32-hex>            # must be `nppt_` + 32 hex chars
NOPPT_ADMIN_KEY=<admin-key>            # required by /api/keys (x-admin-key header)
NOPPT_WEB_URL=http://localhost:5173    # used to build the viewUrl returned to agents
PORT=3001

Option B — issue scoped keys at runtime via the admin API (guarded by the x-admin-key header):

curl -X POST http://localhost:3001/api/keys \
  -H 'Content-Type: application/json' \
  -H 'x-admin-key: <admin-key>' \
  -d '{"name":"hermes-local","tenantId":"hermes","userKey":"local"}'

The plaintext key appears in the response only once — store it immediately. List keys with GET /api/keys and revoke with DELETE /api/keys/:id.

Scopes, artifacts & preview

• Artifacts are written per scope to packages/server/data/tenants/<tenant>/users/<user>/workspace/; different scopes cannot see each other's data (cross-scope reads fail with E4001).

• Agents receive a viewUrl of {NOPPT_WEB_URL}/mcp-preview/{tenant}/{user}/{presentationId}, which renders the deck in a sandboxed, read-only iframe.

• If NOPPT_MCP_VIEW_TOKEN is unset, the view endpoints only allow loopback requests — convenient for local use. Set it (and pass ?token= / X-View-Token) when accessing from another machine.

Agent integration conventions (recommended workflow)

These conventions keep agents (e.g. Hermes) consistent with a "human-in-the-loop" flow. They are not enforced by the server — a cloner can run fine without them, but will generate directly and skip the confirmation page.

• Prefer noppt_prepare_outline_draft; do not use noppt_generate to produce slides directly: the former only stages a draft and returns openUrl, letting the user confirm generation mode / style / slide count / colour theme on the NoPPT Web "AI Generate Presentation" config page before generation; the latter generates asynchronously immediately, bypassing confirmation. Use noppt_generate only when the user explicitly asks for the final deck.

• referenceText budget: 800 × slideCount characters, clamp(3000, 20000); put the most important content first (overflow is truncated from the head). referenceSource is shown on the UI only.

• mode is fixed to auto (unless you intentionally want guided step-by-step).

• Issue your own API key; do not reuse any literal key from outside the repo: use "Way B" admin endpoint to issue your own scoped key, e.g. {"name":"hermes-local","tenantId":"hermes","userKey":"local"}; never commit a local Dev Key (.gitignore already excludes server.env, etc.).

• Four-source RAG (suggested): local wiki/ (L0 single-source LLM Wiki), corpus/ (L1 local corpus), web search (L2), aws-knowledge (L3); conflict arbitration priority L3 ≈ L1 > L2. Note: corpus/ and wiki/ are git-ignored local assets — cloners must supply their own material or pass authoritative text directly via referenceText.

• When to use it: when the user only describes a need without asking for the deck yet, or needs to review the topic / material first, use noppt_prepare_outline_draft; after staging, return only openUrl and wait on the config page for the user's confirmation.

• Ready-to-paste prompt templates: see HERMES_NOPPT_USER_PROMPT.md (per-task user-message version) and HERMES_NOPPT_SYSTEM.md (system-prompt / Skill version) at the repo root. Note: this repo's AGENTS.md is git-ignored (it contains a local key literal); after cloning, issue your own key via the "Issue an API key" section above instead of relying on that file.

🔐 Data Storage & Self-hosting

• Runtime data lives under packages/server/data/ (presentations, drafts, tenants, API keys, logs). Do not commit this directory — it is git-ignored; only config.example.json is tracked as a template.

• Before going public, rotate every API key / secret and replace config.json with config.example.json.

• MCP tool endpoints are protected by bearer API keys (Authorization: Bearer nppt_…); the key management routes (/api/keys) are guarded by the x-admin-key header (NOPPT_ADMIN_KEY).

🤝 Contributing

1. Fork and create a feature branch.

2. pnpm install && pnpm dev:all for local development.

3. Run pnpm lint, pnpm typecheck, and pnpm test before opening a PR.

4. Keep i18n strings in the dictionaries rather than hard-coding user-facing text.

🙏 Acknowledgements

NoPPT was designed by Qiyunlong (齐云龙). Every line of code was built from scratch with AI Coding tools. It draws on the ideas and methods of revealjs-validator, SlidesGen-Bench, and huashu-skills. Special thanks to them.

📄 License

MIT © 2026 NoPPT Authors.