kentjuno/ainovel-cli

ainovel‑cli – AI‑powered multi‑agent novel writer (Vietnamese & multilingual)

What it is – A command‑line / terminal UI (TUI) application written in Go that orchestrates several LLM agents (Coordinator, Architect, Writer, Editor, Arbiter) to automatically draft long‑form fiction. It ships with a fully Vietnamese‑localized interface, supports writing in Vietnamese or Chinese, and can run entirely offline with local Ollama models or connect to cloud providers (OpenRouter, Gemini, Claude, OpenAI, DeepSeek).

Key capabilities

• Multi‑agent workflow – separate agents handle story planning, drafting, editing, and arbitration, with step‑level checkpointing and recovery.
• Two‑tier rolling planning – only the next arc is expanded as the story progresses, keeping context windows manageable for very long novels.
• Prompt caching & 4‑level context compression – reduces token usage while preserving narrative state.
• Rich slash‑command UI – /diag, /simulate, /import, /export, /sync, etc., let you inspect health, mimic a target writing style, import existing texts, and export finished works as TXT or EPUB.
• Model flexibility – you can mix providers per role (e.g., a powerful Claude model for editing and a local Qwen model for drafting) and switch models on the fly.
• Offline first – with Ollama you can run 100 % locally on a GPU (≥12 GB VRAM) using open‑source models such as Qwen 2.5/3.5.
• Vietnamese‑only UI – every menu, status bar, and error message is translated, while the story content can be generated in Vietnamese (default) or Chinese.

Typical workflow

1. Run the setup wizard (first launch) to choose a provider, enter API keys or configure Ollama, and set the language.
2. Start the TUI (docker compose run --rm ainovel or ./ainovel-cli).
3. Either give a one‑sentence premise for a quick start or enter co‑create mode to iteratively refine world‑building and characters.
4. Use slash commands to intervene, import external drafts, simulate a desired author’s style, or export the manuscript.
5. The system saves checkpoints after each tool step, so a crash or Ctrl‑C resumes exactly where it left off.

Installation

• Docker (recommended): docker compose build && docker compose run --rm ainovel.
• From source: Install Go ≥ 1.21, clone the repo, then go build -o ainovel-cli ./cmd/ainovel-cli and run ./ainovel-cli.
• Local LLM: Create an Ollama model with a large context window (e.g., qwen2.5:14b with num_ctx 65536).
• Cloud LLM: Provide API keys for OpenRouter, Gemini, Claude, OpenAI, or DeepSeek in config/config.json.

Directory layout (output lives under workspace/ or a custom NOVEL_DIR):

novels/<title>/output/novel/
├─ chapters/      # finished markdown chapters
├─ drafts/        # intermediate drafts
├─ reviews/       # editor quality reports
├─ summaries/     # arc/volume summaries
├─ premise.md
├─ characters.md
├─ world_rules.md
└─ meta/          # checkpoints, progress, logs

Tech stack

• Language: Go
• Agent core: agentcore (tool‑calling & streaming)
• LLM wrapper: litellm
• TUI framework: Bubble Tea + Lip Gloss
• License: MIT

Who might use it – Writers who want AI assistance for long‑form fiction, hobbyists experimenting with multi‑agent LLM pipelines, or developers looking for a ready‑made example of a sophisticated agent orchestration system.