
Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.
👉 Join on Whop
💬 Join the community for support, free tools and early news on Discord
I built Manuscript Workbench because I kept running into the same problem. Long texts, AI rewrites, copied web pages, Markdown notes, DOCX files and cleanup tasks all ended up scattered across different tools. I wanted one local workspace where I could load text, clean it, structure it, rewrite parts with a local model, preview the result and export to something usable. Without sending everything to a cloud service first.
So that's what this is. One app, running on your own machine, for the entire messy middle part of writing.
Download: https://github.com/cyberdeliaAI/Manuscript-Workbench (free and open source)

What is Manuscript Workbench?
A local Python web app that runs in your browser on 127.0.0.1. No account, no cloud, no telemetry. It's built for long texts: manuscripts, documentation, prompt libraries, articles, stories.
It imports TXT, Markdown, RTF, DOCX, HTML and even URLs. It exports to TXT, Markdown and DOCX. And it talks to local AI through LM Studio, Ollama, oMLX or any OpenAI-compatible endpoint.
Why local-first?
Your drafts stay on your machine. That matters more than you'd think:
Privacy for stories, prompts, model notes and anything else you'd rather not upload
Works with your own local models, including uncensored ones
No subscription, no vendor lock-in
No "we've updated our terms of service" surprises
If you connect AI, it runs through LM Studio, Ollama, oMLX or another OpenAI-compatible local server. That makes it useful for sensitive drafts, private notes, model documentation or long-form creative work.
Main features
1. Load and clean messy text Import TXT, MD, RTF, DOCX and HTML. Extract clean text from web pages, with optional Trafilatura extraction for article content. Repair mojibake, strip invisible control characters and normalize spacing, tabs and line breaks. Copied web text is usually full of junk. This fixes it.
2. Markdown-centered writing Headings with #, ## and ###, live preview, plus quick tools for bold, italic, underline, monospace and small caps. Tab indentation for lists. Everything stays in plain Markdown, so nothing gets locked into a weird proprietary format.
3. Chapter Navigator The app detects your headings and builds a collapsible outline with colored levels. Jump straight to any chapter or section. On a 100k-word manuscript this alone saves your sanity.

4. Local AI rewriting Rewrite a selection, a paragraph, the whole text or chunked long text. There are prompts for conservative cleanup, humanizing text, feedback, questions and translation. Pick your provider, click Detect Model and go.
5. Preview before applying Every rewrite shows a before/after preview. You apply only when you're happy with it. The app also warns when a rewrite would break your document structure. No silent AI mangling.

6. Export TXT for clean plain text, Markdown for the editable manuscript, DOCX for a formatted Word document. Headings, block quotes, scene breaks and inline formatting carry over to Word styles.
Who is this for?
Not just authors. If you're on Civitai, chances are you write more than you think:
AI artists writing model pages and workflow notes
Prompt engineers managing long prompt libraries
Writers working with local LLMs
Anyone cleaning copied web articles or research notes
People converting messy HTML, DOCX or text dumps into Markdown
Anyone writing long Civitai articles, guides or documentation
A typical workflow
Import a messy HTML article or DOCX draft
Extract clean text
Run cleanup checks
Structure the document with Markdown headings
Move through sections with the Chapter Navigator
Rewrite selected paragraphs with a local model
Preview the Markdown
Export to DOCX or Markdown
AI integration
Manuscript Workbench doesn't require one specific AI backend. It's more of a writing interface around your own models than a hosted AI product.
LM Studio: easiest to start with
Ollama: popular and cross-platform
oMLX: great for Apple Silicon and MLX models
Custom endpoint: any OpenAI-compatible server
The app only sends model, messages and stream. Runtime settings like temperature and context length stay in your provider app, unless you fill in the optional Advanced request options.
Why Markdown?
Because it's universal, readable and easy to export. Instead of locking structure into one niche format, headings and sections are stored in a way that works across writing apps, static site generators, GitHub, documentation tools and converters. Your manuscript stays yours, in plain text.
Download and install
Manuscript Workbench is on GitHub:
https://github.com/cyberdeliaAI/Manuscript-Workbench
git clone https://github.com/cyberdeliaAI/Manuscript-Workbench.git
cd Manuscript-Workbench
python -m venv .venv
.venv\Scripts\Activate.ps1 (Windows PowerShell)
# source .venv/bin/activate # macOS/Linux
pip install -r requirements.txt
python app.pyThen open http://127.0.0.1:8765 in your browser. That's it. You only need Python installed, everything else comes from requirements.txt. Want a different port? Set MANUSCRIPT_WORKBENCH_PORT before starting.
The repo also includes the full user guide with every feature explained.
Limitations
Being honest here:
It's not a full word processor replacement
DOCX export is practical, not InDesign-level typesetting
Some websites block URL import with a 403
AI quality depends entirely on your local model
Very large documents can get heavy depending on your machine and model
Feedback welcome
I'm still refining the workflow, especially for long-form AI-assisted writing, local model integration and Markdown/DOCX export. What would make this more useful for your model documentation, prompt collections, stories or Civitai articles? Let me know in the comments.
Manuscript Workbench started as a personal cleanup tool, but it's turning into a local writing environment for people who work with long text and local AI. If you often move between messy drafts, Markdown, DOCX, web text and LLM rewrites, this might save you a lot of friction.
Get it here: https://github.com/cyberdeliaAI/Manuscript-Workbench

