writing-agent 0.4.1

The autonomous long-form writer that argues a thesis and cites real sources - not slop.

One command → a researched, self-critiqued, fact-checked, exported article (or book). Local-first and model-agnostic (any OpenAI-compatible host), on your own key, for cents.

Quickstart · What you get · Examples · Docs ↗

Why it's different

Most AI writing is fluent slop - confident, generic, unsourced. Writing Agent is a pipeline, not a prompt: a durable state machine with a separate critic, an enforced thesis, and claim↔source verification. Every article ships an evidence report - the argument it makes plus every source, ranked by influence. It won't replace a chat when you want to steer every sentence; it replaces the grind of turning a topic into a sourced, non-generic, finished piece. How it works ↗

Try it in 30 seconds

pip install writing-agent                # Python 3.10+ · or: pipx install writing-agent
export OPENROUTER_API_KEY=sk-or-...      # or any OpenAI-compatible host's key - your choice
writing-agent write --abstract "How stoicism applies to modern burnout"

write asks a few questions upfront, then researches, drafts, critiques, verifies, humanizes, and exports the finished file with zero babysitting.

No API key? WRITINGAGENT_FAKE=1 writing-agent write --abstract "test" runs the whole pipeline offline with placeholder output. No terminal? The web demo is a Gradio try-it page you can run from a clone or deploy as a Hugging Face Space. Extras, from-source installs, and host setup: installation guide ↗.

What you get

• Argues, doesn't just cover - a contestable thesis a separate critic enforces in every section, and a side-by-side judge that picks the strongest of N drafts
• Real sources, verified - web-grounded research, and each cited claim is checked against its source; unsupported citations are blocked, not shipped
• Proof, not vibes - the evidence report shows the thesis + every source ranked by influence (0-100), built from the finished piece
• Any field, any voice - registers (fiction → academic → copy), 46 selectable personas, and emotion targets; the craft engine holds up even on cheap models
• Self-correcting, optionally self-directing - fixed quality gates by default; an opt-in agentic controller that chooses its own next move when you want it
• Cheap when you want it - cost_mode: budget targets a finished article for ~$0.10-0.25 depending on the model
• Durable - plain markdown + JSON on disk; kill a run mid-flight and run resumes exactly where it stopped
• Ships, not just writes - 6 export formats (pdf · epub · html · docx · txt · md), self-laying-out diagrams, free-licensed or AI-generated images, an on-page SEO audit, a promo pack (X thread, LinkedIn, headlines), and one-paste Copy for Medium / Substack / X

Four ways to use it

There's also an embeddable Python API - see docs ↗.

See real output

Judge it before you install it: examples/ has finished articles with their evidence reports, an A/B pilot against a plain chatbot, and examples/sample-run/ - a complete autonomous book run with every working file the agent produced.

Learn more

Everything else lives in the docs ↗: guides (articles, books, research, exporting), the full command and settings reference, model routing, cost & performance, and the FAQ. The engineering spec is docs/plan.md; the UI system is docs/design.md.

Contributing

git clone https://github.com/vikast908/WritingAgent && cd WritingAgent
pip install -e ".[dev]"
ruff check . && pytest        # the suite runs fully offline - no API key

Issues and PRs welcome - see CONTRIBUTING.md, plus SECURITY.md and the CHANGELOG.

Author

Built and maintained by @vikast908. Questions, ideas, or bugs? Open an issue.

License

MIT.