What it does

WriteHERE is a Python framework for long-form writing—fiction or technical reports—that breaks tasks down recursively and interleaves planning with execution. It runs as a backend engine or with a React frontend that visualizes the agent’s “thinking process” in real time. You feed it a prompt file, pick a model (GPT-4o, Claude 3 Sonnet, etc.), and choose story or report mode.

The interesting bit

Most AI writing tools lock you into a fixed outline-then-draft pipeline. WriteHERE’s engine (recursive/engine.py) decomposes tasks on the fly, mixing retrieval, reasoning, and composition steps as needed. The frontend renders this as a live task graph—handy for watching an agent wander down a wrong branch before correcting itself.

Key highlights

• Supports both creative fiction and technical report generation, with search integration via SerpAPI for the latter
• Real-time web visualization of hierarchical task decomposition and status
• Transparent operation: all agent decisions and prompts are exposed, not black-boxed
• One-script setup (./start.sh) spins up backend (port 5001) and frontend (port 3000)
• EMNLP 2025 oral paper backing the approach; MIT licensed and explicitly non-commercial

Caveats

• Requires API keys for OpenAI, Anthropic, and SerpAPI; no local model support mentioned
• Python 3.6+ and Node.js 14+ prerequisites, plus dependency wrangling that merits a dedicated Anaconda script
• The README claims it “consistently outperforms state-of-the-art methods” but offers no actual metrics or benchmark names

Verdict

Worth a look if you’re researching agentic writing workflows or want a transparent, hackable alternative to closed AI writing tools. Skip it if you need offline-only operation or are allergic to API costs and Node.js frontend setups.