Installation guide

Full local setup

Everything local, at your own pace. Half of Palimpseste works the moment you install it, with nothing else. The other half — AI, Zotero, some exports — activates when you add the third-party tools you choose. This guide gives you the full picture.

Download Palimpseste Start the guide

1. What works the moment you install

No external dependencies. You download the DMG, EXE or AppImage, and everything below works immediately, offline.

  • ✓ Four editing modes — Raw Markdown, Preview, Split, WYSIWYG (TipTap)
  • ✓ Offline spellchecker in French, English, German, Spanish, Italian (embedded Hunspell, ~25 MB)
  • ✓ Native UI in five languages — 2,287 keys × 5 languages, strict parity
  • ✓ Novel mode in manual editing — Story Bible, plot threads, drag-and-drop corkboard, timeline, synopsis. All cards (characters, locations, factions, scenes, narrative arcs, synopses) are created and edited by hand, without any AI dependency. AI regeneration of these panels is an optional add-on (see § 2).
  • ✓ Researcher mode in manual editing — thesis project, academic glossary, compliance, coherence checks (without Zotero or LaTeX). All editing works without AI.
  • ✓ Journalist mode in manual editing — 5 article templates, deadline tracker, sources with 3 confidentiality levels, AP/Le Monde/AFP compliance. All editing works without AI.
  • ✓ Inline CSL citations — 5 styles (APA, Chicago, MLA, IEEE, Harvard) without Zotero
  • ✓ Native export — Markdown, HTML, LaTeX, EPUB, DOCX, JATS XML, manuscript format, presentation. Native JavaScript converters, no external tool needed.
  • ✓ Native import — DOCX, HTML, RTF, Markdown, plain text via JavaScript converters
  • ✓ Semantic RAG on your documents — fastembed-rs ONNX model downloaded once (~480 MB) on first launch, then 100% local. No cloud call for indexing.
  • ✓ Semantic versioning — named snapshots, word-level diff, supervision mode
  • ✓ Personas and My Studio — local library of reusable author voices
  • ✓ Markdown extensions — Mermaid, KaTeX, footnotes, wiki-links, Zettelkasten backlinks
  • ✓ Style analysis — repetition detection, language tics, personal style dictionary
  • ✓ Basic PDF — via the system's print dialog (no LaTeX needed)

2. To use generative AI

Everything that requires generative AI — AI chat, five verbs (rephrase, continue, summarize, critique, translate), AI enrichment of the Story Bible, AI detection of plot threads, AI generation of synopses, Story Engine — goes through an AI provider you choose. Four paths, combinable. Note: novelist / researcher / journalist panels fully work in manual editing without any AI (see § 1); AI is only an optional augmentation layer.

2.a — Ollama (recommended for most)

100% local · Free

The simplest way to run an AI model locally. Clean CLI, auto-start, dynamic memory management. OpenAI-API compatible, so Palimpseste auto-detects it on localhost:11434.

  1. Download Ollama from ollama.com (macOS / Windows / Linux)
  2. Run the installer, accept the permissions
  3. Open a terminal and run: 
    ollama pull qwen2.5:32b
  4. In Palimpseste → Preferences → AI, select Ollama. The model appears automatically.

Recommended models per your RAM (see also § 4): Qwen 2.5 32B (32 GB RAM, excellent quality/performance ratio), Gemma 3 27B (32 GB), Mistral Small 3 (16 GB, faster), Phi-4 14B (16 GB, very good for short tasks), Llama 3.3 70B (64 GB, maximum quality).

2.b — LM Studio (graphical UI, more accessible)

100% local · Free

A graphical app to discover, download and run GGUF models in a few clicks. More visual than Ollama, ideal if you prefer to avoid the command line. Exposes an OpenAI-compatible server on localhost:1234.

  1. Download LM Studio from lmstudio.ai (macOS / Windows / Linux)
  2. Launch the app and, in the "Discover" tab, search for a model (for example Qwen2.5-32B-Instruct, Gemma-3-27B, or Mistral-Small-3.1)
  3. Download the model (several GB depending on size), then load it into memory in the "Chat" tab
  4. Start the local server from the "Local Server" tab (port 1234 by default)
  5. In Palimpseste → Preferences → AI, select LM Studio. The loaded model appears automatically.

2.c — Cloud (Anthropic Claude or OpenAI GPT)

Cloud · Your key, your cost

For machines without enough RAM for local, or simply to benefit from the best-in-class models. You provide your own API key, your requests go directly from Palimpseste to the provider — no middleman markup, no cross-subscription. You stay in control of the cost via your billing dashboard with the provider.

  1. Create an account at console.anthropic.com or platform.openai.com
  2. Generate an API key (Anthropic: "API Keys"; OpenAI: "API keys")
  3. Set a spending limit (recommended) in your provider account settings
  4. In Palimpseste → Preferences → AI, select Anthropic or OpenAI and paste the key. It is stored in your system's secure keychain (Keychain macOS, Credential Manager Windows, libsecret Linux).

Cost ranges: on Anthropic Claude Sonnet 4.6, an intensive Story Engine session (5 to 10 chapter generations) typically costs between €0.50 and €2. On OpenAI GPT-4o, expect 30 to 50% more for equivalent quality.

2.d — OpenAI-compatible gateway (LiteLLM, vLLM…)

Advanced · Since v1.15

Any server speaking the OpenAI API becomes a first-class provider: a LiteLLM gateway federating several providers, a team vLLM server, a corporate proxy, a remote machine on your VPN. In Settings → AI Assistant, pick "OpenAI-compatible", enter the server URL (forms with or without a trailing /v1 both work), the API key if your gateway requires one (stored in the system keychain), then Test connection: the list of actually served models fills itself — just pick one.

3. Optional integrations

For specific use cases, you can activate three additional integrations. None of them are required for everyday use.

3.a — Zotero (BBT citations or Cloud)

Researcher profile · Recommended

For researchers who manage their bibliography in Zotero, Palimpseste connects in two ways. Better BibTeX (local, recommended): a Zotero plugin that exposes a local HTTP API; your citations appear in the Cmd+K menu. Zotero Cloud (online): Zotero API key, simpler to configure but requires a connection. Without Zotero, you can still use inline CSL citations directly in markdown (5 styles).

  1. Install Zotero from zotero.org (free, open source)
  2. For BBT: install the Better BibTeX plugin in Zotero (also free)
  3. Launch Zotero (BBT auto-starts with the app)
  4. In Palimpseste → Preferences → Citations, choose "Better BibTeX (local)" or "Zotero Cloud" and paste the API key if applicable

3.b — TeX Live or MacTeX (academic PDF with LaTeX)

Optional · For high-quality PDF

Palimpseste natively exports to LaTeX (a .tex file ready to compile) with no dependency. To compile this .tex to PDF — professional academic typography, perfectly rendered equations, thesis layout — you need a LaTeX distribution on your machine. Not necessary if a basic PDF via the system's print dialog is enough for you.

  1. macOS: install MacTeX (~5 GB). Linux: sudo apt install texlive-full or equivalent. Windows: TeX Live.
  2. In Palimpseste, export to LaTeX, then in a terminal: pdflatex myfile.tex or xelatex for accented/Unicode characters

3.c — Pandoc (exotic formats only)

Optional · Marginal cases

Palimpseste has its own native JavaScript converters for common formats: DOCX, HTML, RTF, LaTeX, EPUB, JATS XML work out of the box. Pandoc is only useful for less common formats: ODT (LibreOffice), Org-mode (Emacs), ConTeXt, RST (reStructuredText) and a few others. If you don't use any of these formats, you can skip Pandoc — most users don't need it.

  1. Install Pandoc from pandoc.org/installing.html (free, open source)
  2. Palimpseste auto-detects it and activates additional formats in the Export menu

4. Hardware recommendations for local AI

Local AI is the most demanding component. Here is what your machine can realistically pull off.

Setup Suitable models Use case
8 Go RAM ou moins Cloud only, or very small 3B models (Phi-3-mini, Gemma 2B) Local AI works but quality is low. Prefer Anthropic / OpenAI.
16 Go RAM Phi-4 14B, Mistral Small 3.1, Llama 3.2 11B Good for short AI verbs, chat, structured output extraction. Story Engine works, acceptable quality.
32 Go RAM ★ Qwen 2.5 32B, Gemma 3 27B, Mistral Small 3.1, Phi-4 14B Sweet spot. Excellent quality for Story Engine, convincing chapter generation, stable structured output. Apple Silicon M2+ favored.
64 Go RAM ou plus Llama 3.3 70B, Qwen 2.5 72B, MoE models like Qwen3 235B Cloud-class quality, full autonomy. Apple Silicon M3 Max / M4 Max or GPU 24 GB+.

Apple Silicon note. Unified memory on M-series Macs is particularly efficient for local AI: models run at the same speed as on dedicated GPU with less total RAM. A 32 GB MacBook Pro M3 outperforms most 32 GB PCs on these tasks.

5. Summary — what requires what

Capability Included? If not, depends on…
4-mode Markdown editor ✓ —
Story Bible, plot threads, corkboard, timeline (manual editing) ✓ —
AI regeneration of Story Bible, plot threads, synopses — Ollama, LM Studio, Anthropic or OpenAI (see § 2)
5-language spellchecker ✓ —
Markdown / HTML / LaTeX / EPUB / DOCX / JATS export ✓ —
Inline CSL citations (5 styles) ✓ —
Local semantic RAG ✓ (~480 Mo téléchargés au 1er lancement) —
AI chat, 5 contextual verbs (rephrase, continue…) — Ollama, LM Studio, Anthropic, or OpenAI
Story Engine (chapter generation) — Same as above (32 GB RAM recommended for local)
Zotero citations (full bibliography) — Zotero + Better BibTeX, or Zotero Cloud API key
Academic PDF via LaTeX — TeX Live or MacTeX
Basic PDF (system print) ✓ —
ODT, Org-mode, ConTeXt export — Pandoc

Ready to set up your workshop?

Download Palimpseste, start with nothing else, and add integrations at your own pace — or never.

Download Palimpseste See Story Engine