Local-first, autonomous agent operating system for personal AI assistance.
Cognition + Thor — Intelligence with Power
Ollama, LM Studio (both local), OpenAI, Anthropic, Gemini, Groq, DeepSeek, Mistral, Together, OpenRouter, xAI, Cerebras, GitHub Models, AWS Bedrock, Hugging Face, Moonshot. Auto-detects from API key.
CLI, Web UI, REST API, Telegram (poll + webhook), Discord, Slack, WhatsApp, Signal, iMessage, Teams, Matrix, Google Chat, Mattermost, Feishu, IRC, Twitch, Voice (STT/TTS with wake word).
Core identity, episodic logs, semantic knowledge graph, procedural skills, working memory. 3-channel hybrid search: BM25 + vector embeddings + graph traversal.
Deterministic Gatekeeper (4 risk levels), 4-level sandbox, SHA-256 audit chain, Fernet AES-256 token encryption, API rate limiting, CORS hardening, XSS prevention. 80-item audit with 0 remaining.
Full dashboard (React 19 + Vite 7) with integrated chat, voice mode, live config editing, agent management, prompt editing, cron jobs, MCP servers, A2A settings. Pre-built UI ships with the installer.
10 tool modules: filesystem, shell, memory, web, browser, media, vault, synthesis, code, skills. A2A protocol (Linux Foundation RC v1.0) with Planner delegation — agents can discover and delegate tasks to remote agents.
Planner (LLM) → Gatekeeper (deterministic policy) → Executor (DAG-based parallel sandbox). No hallucinated tool calls — every action is policy-checked before execution.
Wake word ("Jarvis") with Levenshtein + phonetic matching, conversation mode, Piper TTS (Thorsten Emotional), Whisper STT, integrated chat with WebSocket streaming.
Obsidian-compatible Markdown vault with YAML frontmatter, tags, backlinks, full-text search. Knowledge synthesis with confidence ratings, contradiction detection, and gap analysis.
All data stays on your machine. No cloud required — Ollama runs fully local. Optional cloud LLM providers available. Full GDPR compliance toolkit.
Personality Engine (warmth, humor, greetings), sentiment detection (frustrated/urgent/confused/positive), user preference learning, real-time status callbacks, multilingual error messages.
Reflector auto-synthesizes reusable skills from successful sessions. Learned procedures are suggested for future similar requests. Skill marketplace with ratings and remote registry.
Install, search, rate, and report community skills from a GitHub-hosted registry. Publisher verification with 4 trust levels. 5-check validation pipeline. ToolEnforcer runtime sandboxing.
Full internationalization with English, German, and Chinese (Simplified). 244 translated keys per locale, SHA-256 integrity verification, curated Planner system prompts. Add new languages with a single JSON file.
┌───────────────────────────────────────────────────────────────┐ │ Control Center UI (React 19 + Vite 7) │ │ Config · Agents · Chat · Voice · Prompts · Cron · MCP · A2A │ ├───────────────────────────────────────────────────────────────┤ │ REST API (FastAPI, 20+ endpoints, port 8741) │ ├───────────────────────────────────────────────────────────────┤ │ Channels (17) │ │ CLI · Web · Telegram · Discord · Slack · WhatsApp · Voice │ ├───────────────────────────────────────────────────────────────┤ │ Gateway Layer │ │ Session · Agent Loop · Personality · Sentiment · Prefs │ ├───────────────────────────────────────────────────────────────┤ │ Context Pipeline (Memory · Vault · Episodes) │ ├────────────┬─────────────┬────────────────────────────────────┤ │ Planner │ Gatekeeper │ Executor (DAG parallel) │ │ (LLM) │ (Policy) │ (Sandbox) │ ├────────────┴─────────────┴────────────────────────────────────┤ │ MCP Tool Layer (53 tools, 10 modules) │ │ Filesystem · Shell · Memory · Web · Browser · Media │ │ Vault · Synthesis · Code · Skills Marketplace │ ├───────────────────────────────────────────────────────────────┤ │ Multi-LLM Backend Layer (16) │ ├───────────────────────────────────────────────────────────────┤ │ 5-Tier Cognitive Memory │ │ Core · Episodic · Semantic · Procedural · Working │ └───────────────────────────────────────────────────────────────┘
From clone to running agent in under 5 minutes.
# Option A: One-Click (Windows) — no Node.js needed
# Double-click start_cognithor.bat -> Browser opens -> Power On -> Done.
# Auto-installs Python + Ollama via winget if missing.
# Option B: Manual install
git clone https://github.com/Alex8791-cyber/cognithor.git
cd cognithor
pip install -e ".[all,dev]"
# Pull Ollama models
ollama pull qwen3:32b # Planner (20 GB VRAM)
ollama pull qwen3:8b # Executor (6 GB VRAM)
ollama pull nomic-embed-text # Embeddings
# Start
cognithor # Interactive CLI
cognithor --lite # Lite mode (6 GB VRAM total)No GPU? Use cognithor --lite or set a cloud API key. See LLM Providers for all 16 options.