tinyhumansai/openhuman

What it solves

OpenHuman provides a local-first AI assistant that eliminates the "cold start" problem where agents lack personal context. It replaces fragmented AI tools with a unified orchestrator that manages memory, multi-agent workflows, and deep research across a user's entire digital stack.

How it works

The system functions as a three-part architecture:

• The Brain: It uses "Memory Trees" to compress user data from 100+ OAuth integrations into scored Markdown files stored in a local SQLite database and mirrored as an Obsidian vault. An auto-fetch loop updates this context every 20 minutes.
• The Orchestrator: Instead of simple loops, it runs agents on checkpointed graphs (via tinyagents) that can pause for human approval and survive restarts. It supports encrypted agent-to-agent communication using the Signal protocol.
• The Researcher/Doer: It integrates native tools for web search (Exa), scraping, browsing, voice (Whisper), and media generation, while routing workloads to the most appropriate LLM (local via Ollama or cloud providers).

Who it’s for

Individuals who want a highly personalized, private AI agent that has full context of their emails, documents, and messages without relying on cloud-only proprietary systems.

Highlights

• Local-First Memory: Persistent memory stored as an editable Obsidian wiki rather than a black-box vector database.
• Visual Workflows: Agents propose automation graphs that users can review and edit on a visual canvas.
• Massive Integration Library: Supports over 100 OAuth integrations, 5,000+ MCP servers, and 90,000+ skills.
• Privacy Mode: A hardware-enforced switch that ensures no inference leaves the local machine.
• Token Compression: Uses TokenJuice to reduce tool output tokens by up to 80% to lower costs.
• Multi-Channel Reach: Connects to 17 messaging channels, including native email (IMAP/SMTP).