Identity Preserving Publishing Engine: a CLI that builds voice, knowledge, and publication profiles from your writing, then generates articles that sound like you.
# 1. Initialize
thoth init
# 2. Import your writing
thoth import voice ~/notes ~/essays
thoth import knowledge ~/architecture ~/repos
thoth import publications ~/articles
# 3. Generate identity profiles
thoth generate_profile
thoth profile_status
# 4. Research a topic
thoth research "offline first mobile architecture"
# 5. Generate an article in your voice
thoth generate_article --topic "offline first mobile architecture"
# 6. Export
thoth export_article <id>npm install -g @serticode/thoth- Node.js 22+
- C++ build tools —
better-sqlite3andsqlite-vecare native C++ modules:- macOS: Xcode CLI tools (
xcode-select --install) - Linux:
build-essential(or equivalent) - Windows: Visual Studio Build Tools with C++ workload
- macOS: Xcode CLI tools (
If you don't have a C++ compiler, the install will fail with a node-gyp error. Install your platform's build tools and try again.
ollama pull qwen2.5:7b
ollama pull nomic-embed-textWith Ollama running locally, Thoth works with zero API keys.
API keys are read from environment variables. Set at least one:
export OPENAI_API_KEY=sk-... # best quality, tried first
export GROQ_API_KEY=gsk_... # fast inference
export GEMINI_API_KEY=AIza... # free tier available
export ANTHROPIC_API_KEY=sk-ant-... # Claude modelsNo keys? Ollama is always available as a local fallback.
All models are overridable:
export OPENAI_CHAT_MODEL=gpt-4o
export OLLAMA_CHAT_MODEL=qwen2.5:7b
export OLLAMA_EMBEDDING_MODEL=nomic-embed-textSee thoth.json for project level overrides and the full config docs.
| Command | Description |
|---|---|
thoth init |
Create database and run migrations |
thoth import voice <path> |
Import voice sources (journals, essays, notes) |
thoth import knowledge <path> |
Import knowledge sources (architecture, repos, docs) |
thoth import publications <path> |
Import publication sources (articles, blog posts) |
thoth generate_profile |
Generate all identity profiles via AI |
thoth profile_status |
Show ready/missing status for each profile |
thoth research "<topic>" |
Research a topic with vector-search + AI synthesis |
thoth generate_article --topic "<topic>" |
Generate an article in your voice |
thoth list_articles |
List all generated articles |
thoth get_article <id> |
Show article content |
thoth create_series <name> |
Create a new article series |
thoth list_series |
List all series |
thoth add_to_series <series> <article> |
Add an article to a series |
thoth get_series <id> |
Show series with all articles |
thoth export_article <id> |
Export article (md, html, txt, rss) |
thoth export_series <id> |
Export entire series |
Output goes to ~/Documents/Thoth/singles/ and ~/Documents/Thoth/series/.
Thoth chains through AI providers in quality priority order. Each provider is gated by its API key; if the key isn't set, it's skipped.
| Provider | Chat Model | Embedding Model | Key Required |
|---|---|---|---|
| OpenAI | gpt-4o-mini |
text-embedding-3-small |
OPENAI_API_KEY |
| Groq | mixtral-8x7b-32768 |
— | GROQ_API_KEY |
| Gemini | gemini-1.5-flash |
text-embedding-004 |
GEMINI_API_KEY |
| Anthropic | claude-3-haiku |
— | ANTHROPIC_API_KEY |
| Ollama | qwen2.5:7b |
nomic-embed-text |
none (local) |
Each provider has a 15 second timeout. If one fails, the next is tried automatically.
src/
├── domain/ # Pure TypeScript — entities, repository interfaces, Result<T>
├── application/ # Use cases — import, generate profiles, research, articles, series
├── infrastructure/ # Adapters — SQLite, OpenAI SDK, file system, logging
└── cli/ # Commander.js — thin delegation to use cases
Clean Architecture with strict layer separation. Domain has zero external dependencies. Infrastructure implements domain interfaces. CLI depends on application use cases.
All data stays local. Voice profiles, knowledge profiles, embeddings, research notes, and generated articles are stored in ~/.thoth/thoth.db (SQLite). Nothing is sent to any server except the AI provider you configure. Medium publishing is draft-only and requires explicit user action.