Personal AI Sovereignty
Thoth is a local-first AI assistant with 27 core tools, Designer Studio with five modes (decks, documents, landing pages, app mockups, storyboards), a personal knowledge graph, 5 messaging channels, voice, vision, shell access, browser automation, advanced workflows, image and video generation, in-app auto-updates, self-aware status tooling, and a plugin marketplace. Run everything locally via Ollama, or add cloud models when you need frontier reasoning. Either way, your projects, memories, and conversations stay on your machine.
Governments invest billions to keep AI infrastructure within their borders. Thoth applies the same idea to you - your compute, your data, your choice of model, accountable to no one but you.
Run local for full privacy, add cloud when you need it, or both at once
Every LLM call runs on your hardware. No data leaves your machine, ever.
qwen3:14b (~9 GB, runs well on 16 GB RAM)No GPU required. Only the current conversation is sent - memories, files, and history never leave.
Bottom line: Cloud AI assistants rent you access to someone else's system. Thoth gives you ownership - run local, add cloud, or both, and keep your data on your machine either way. Start with cloud today, and as local models get smarter and hardware gets cheaper, transition to fully local, fully private, fully free AI - seamlessly.
Thoth builds a connected web of people, places, preferences, events, and their relationships - not just isolated facts. 10 entity types linked through 67 typed directional relations with 60+ aliases and automatic normalisation, so the agent can reason about how things in your life connect.
Graph-enhanced auto-recall surfaces relevant memories plus their connections before every response. Vague relations are automatically rejected; a curated vocabulary keeps the graph clean. An interactive network diagram renders everything visually - search, filter by type, click nodes for details, toggle between full-graph and ego-graph views.
Reasons about your question, decides which of 27 core tools to use, chains multiple lookups, and assembles a complete answer - all autonomously. Dynamic identity and self-knowledge context keep responses grounded in current capabilities, while smart trimming keeps multi-tool chains within window limits.
Run shell commands through natural conversation - install packages, manage git repos, run scripts. 3-tier safety: safe commands run instantly, moderate commands require approval, dangerous commands are blocked. Persistent sessions per thread.
Autonomous web browsing in a real Chromium window - navigate, click, fill forms, scroll, manage tabs through conversation. Accessibility-tree snapshots let the agent see and interact with elements by number. Persistent profile keeps you logged in.
Local speech-to-text via faster-whisper (CPU, int8 quantized) and neural TTS via Kokoro with 10 voice options. Mic auto-mutes during playback. Fully hands-free, fully offline - no cloud APIs touch your voice data.
Step-based pipelines with conditional branching, approval gates, and 7 schedule types. 5 step types - Prompt, Condition, Approval, Subtask, Notify - with variable passing between steps. Workflow Console now surfaces live progress, approvals, quick launch, recent runs, and actionable insight cards.
Upgraded in 3.16Track medications, symptoms, exercise, periods, mood, sleep - through natural conversation. “I took my Lexapro” → logged. 7 built-in analyses: adherence, streaks, trends, co-occurrence. Auto-export to CSV and interactive Plotly charts.
Build in five modes - decks, documents, landing pages, app mockups, and storyboards. Live editor with runtime preview bridge, surgical AI edits with mutation diff/review, brand lint, reusable components, AI image and video generation, charts, presenter mode, published links, and isolated PPTX export. Home-screen gallery, unified setup flow, and a thread-isolated working surface keep large designs responsive.
Upgraded in 3.17Thoth checks for new releases on a quiet schedule, verifies SHA-256 hashes and code signatures (signtool on Windows, codesign on macOS) before launching the installer, and surfaces a status-bar pill plus a What's New dialog. Stable and beta channels, debounced reminders, and a one-click skip - all wired through a stdlib-only daemon.
New in 3.17Query version, model, tools, channels, logs, errors, Designer state, and more through the built-in Thoth Status tool. Configure assistant name and personality, inspect live system state, and apply selected self-management actions without leaving the app.
Upgraded in 3.175 channels - Telegram, WhatsApp, Discord, Slack, and SMS - with full parity: streaming responses, typing indicators, reactions, media support, slash commands, and approval routing. Tunnel manager auto-exposes webhook ports. Shared media capture, Discord voice fixes, and a live channel monitor in the sidebar with status dots and last-activity tracking.
Upgraded in 3.17Capture from webcam or take screenshots and ask questions about them. Works with local vision models or cloud models. Multi-camera support, inline image display in chat.
Search emails, read threads, draft and send messages with attachments. View, create, update, and manage calendar events - all through natural language. Tiered permissions let you control read, write, and destructive access independently.
Sandboxed file operations with PDF, CSV, Excel, JSON support. Structured files return schema + stats + preview via pandas. Interactive Plotly charts rendered inline - bar, line, scatter, pie, heatmap, and more.
Generate and edit images through conversation via OpenAI, xAI (Grok Imagine), and Google (Imagen 4, Gemini). Generate short videos with Google Veo. Describe what you want, the agent renders it inline. Paste or attach an existing image to edit - per-provider model picker shows only models for configured providers.
Upgraded in 3.17Browse a GitHub-hosted marketplace, install with one click, manage from Settings. Plugins run in a security sandbox with pattern blocking, dependency safety, and 5-second load timeouts. Hot-reload on enable/disable. Build your own with the PluginAPI.
New in 3.1212 reusable instruction packs plus 17 tool guides - Data Analyst, Deep Research, Design Creator, Knowledge Base, Meeting Notes, Daily Briefing, and more. Manual skills toggle from Settings; tool guides auto-activate when their linked tools are enabled.
A personal wiki the agent reads and writes - Markdown articles organized by topic, cross-linked with the knowledge graph. FAISS-indexed for semantic retrieval alongside your documents.
A 5-phase nightly process that autonomously enriches your knowledge graph - discovers missed entities, infers new relationships, decays stale connections, prunes orphans, and turns recent system activity into actionable insights. Hub diversity caps, batch rotation, rejection caching, and Ollama busy checks keep it efficient and accurate.
Upgraded in 3.16Automatically extracts entities and relationships from uploaded documents into the knowledge graph. PDF, DOCX, TXT, and CSV analyzed on upload. Curated relation vocabulary, entity caps, description quality gates, and self-loop rejection keep the graph clean. Provenance tracking links every extracted fact back to its source.
Upgraded in 3.13One-click installer, native OS window via pywebview, system tray with status indicator, branded splash screen, and a first-launch setup wizard. No Docker, no YAML, no terminal needed.
Thoth maintains a structured self-knowledge manifest of its tools, channels, models, and design capabilities, then turns Dream Cycle observations into actionable insights. Workflow Console actions let you pin, dismiss, investigate, and apply suggested fixes or skill proposals from one place.
New in 3.1639 curated local models plus cloud models from OpenAI, Anthropic, Google, xAI, and OpenRouter. Full reasoning model support - thinking tokens in a collapsible bubble with persistence across reloads. Per-thread and per-workflow model switching, star your favourites. Configurable context windows up to 256K tokens.
Ask a question in natural language and the agent autonomously decides which tools to chain together. Search the web, check your email, run a shell command, inspect Thoth's own status, build a deck in Designer, and chart the results - all in a single conversation turn.
Every core tool works locally by default. Connected channels inject their own send tools automatically, while external APIs remain opt-in and individually disableable.
The agent picks the right core tool, then auto-generates send tools for any connected channel
Build presentations and one-pagers in Designer, automate recurring work with step-based pipelines, and route results or approval requests to any of 5 messaging channels. Extend the agent with community plugins from the built-in marketplace - or write your own.
Designer projects stay local, plugins run sandboxed with pattern blocking and dependency safety, and the channel framework auto-generates tools for each connected platform.
Cloud AI rents you access. Thoth gives you ownership.
| Cloud AI | 𓁟 Thoth | |
|---|---|---|
| Your data | Stays on your machine - always. Cloud models send only the current conversation | |
| Conversations | Local SQLite, fully yours, exportable anytime | |
| Cost | Free with local models. Cloud: pennies per conversation | |
| Memory | Knowledge graph - 10 entity types, 67 relation types, visual explorer, FAISS semantic search | |
| Tools | 27 core tools + channel send tools - Gmail, Calendar, shell, browser, vision, Designer, Thoth Status, X | |
| Customisation | Swap models per thread, mix local + cloud, configure identity, enable skills and guides, build workflows, install plugins | |
| Voice | Local Whisper STT + Kokoro TTS - never leaves your mic | |
| Messaging | 5 channels - Telegram, WhatsApp, Discord, Slack, SMS - with streaming and media | |
| Availability | Local models work fully offline |
One-click Windows installer or macOS DMG - both set up Python and all dependencies automatically. Or clone the repo and run from source in 5 commands.
First-launch wizard: Local - download an Ollama model (default: qwen3:14b), or Cloud - enter your API key. No GPU required for cloud.
Ask anything in natural language - the agent handles the rest. Search the web, check email, analyze your screen, track a medication, schedule a workflow.
Thoth doesn't promise privacy - it enforces it structurally. Voice, memory, conversations, and knowledge graphs run entirely on your machine. LLM inference is local by default; opt-in cloud models send only the current conversation. External network calls happen only when you explicitly use an online tool - and each one can be individually disabled.
Free, open source, and yours to keep. No accounts. No subscriptions. No data harvesting.