OpenHuman

OpenHuman is a local-first desktop assistant that combines a persistent memory of your own data, an agent orchestrator, and a research and tool layer in one installable app. It is distributed as desktop installers and is in early beta.

Memory is stored as scored Markdown trees in SQLite on your machine and mirrored as an Obsidian vault you can open and edit directly. Auto-fetch pulls from connected accounts every 20 minutes, so mail, calendar, repositories, documents, and messages are compressed into that memory rather than gathered per conversation. Connectivity covers OAuth integrations, MCP servers, and skills, and tool output passes through a compression layer the project calls TokenJuice before it reaches the model. An optional memory backend proxies to agentmemory when you already self-host it.

Turns run as checkpointed graphs rather than a single loop, so they pause for a human, survive a restart, and resume mid-run. Sub-agent fleets spawn up to three levels deep, stuck agents produce root-cause reports, and instances message each other over Signal-protocol end-to-end encrypted sessions with x402 payments. Workflows are proposed by the agent, reviewed on a visual canvas, and then run durably on schedules, webhooks, or channel events with approvals gating side effects.

The tool layer includes managed web search through Exa, a scraper, a coder toolset, a browser and computer-use surface, in-process Whisper voice, and image and video generation. Model routing selects a model per workload; you can use the project's subscription, your own provider key, or a local Ollama model, and Privacy Mode blocks all inference from leaving the machine, enforced in the Rust core. The agent is reachable through 17 messaging channels including Telegram, Discord, Slack, WhatsApp, Signal, iMessage, and native email over IMAP IDLE and SMTP.

Features

• Memory Tree: your data compressed into scored Markdown trees in local SQLite, mirrored as an editable Obsidian vault
• Auto-fetch: connected accounts sync into memory on a 20-minute loop
• Integrations: OAuth connectors, MCP servers, and skills, plus an optional agentmemory backend
• Token compression: tool output is compressed before it reaches the model
• Graph runs: checkpointed turns that pause for a human, survive restarts, and resume mid-run
• Sub-agent fleets: specialists spawn three levels deep, with root-cause reports for halted agents
• Encrypted agent-to-agent: instances orchestrate each other over Signal-protocol sessions with x402 payments
• Visual workflows: the agent proposes an automation, you review it on a canvas, and it runs on triggers with approval gates
• Native tools: web search through Exa, scraper, coder toolset, browser and computer use, voice with in-process Whisper, image and video generation
• Model routing: per-workload model selection across the bundled subscription, your own keys, and local Ollama models
• Privacy Mode: a single switch stops any inference leaving the machine, enforced in the Rust core
• 17 channels: Telegram, Discord, Slack, WhatsApp, Signal, iMessage, and native email over IMAP and SMTP