Skip to content

OpenHuman (TinyHumans AI)

Overview

OpenHuman is an open-source (GPL-3.0) desktop personal AI agent developed by TinyHumans AI, launched in beta on May 13, 2026. It positions itself as a "personal AI super intelligence" — a local-first agent that ingests data from 118+ connected services, builds a persistent 1-billion-token memory of the user's life and work, and acts autonomously on their behalf. The project hit GitHub trending on its first day (+1,694 stars/day) and reached 27k+ stars within two weeks of launch.

Unlike task-scoped coding agents, OpenHuman is designed as a long-lived companion: it continuously absorbs context from email, calendars, code repositories, and messaging platforms, then uses that context to answer questions, complete tasks, and participate in video meetings as a speaking avatar.

Key Capabilities

NeoCortex Memory Layer

The central differentiator is NeoCortex — TinyHumans' proprietary persistent memory system:

  • Processes up to 10 million tokens at 4,000 tokens/second to build a living knowledge graph
  • Supports up to 1 billion tokens of cumulative cross-session memory
  • Data is canonicalized into ≤3k-token Markdown chunks, scored for importance, and folded into hierarchical summary trees
  • All data stored in SQLite on the local machine — not sent to a cloud data store
  • Auto-fetch pulls fresh data from connected services every 20 minutes
  • Storage format is Obsidian-compatible: the memory vault opens directly in Obsidian for inspection and editing

Third-Party Integrations (118+)

One-click OAuth connections managed via the Composio connector layer:

Category Representative Services
Email & Calendar Gmail, Google Calendar
Storage & Docs Google Drive, Notion, Dropbox
Code & Dev GitHub, Linear, Jira
Communication Slack
Finance Stripe
And many more 110+ additional services

The integration pipeline automatically ingests new data and folds it into the Memory Tree without manual prompting.

Agent Toolbelt

Built-in tools available to the agent for task execution:

Group Tools
Web Web search, web scraping
Coding Filesystem read/write, git, lint, test runner, grep
Voice Speech-to-text (STT) input, ElevenLabs TTS output
Meetings Live Google Meet agent (see Mascot below)

Mascot and Voice Interface

OpenHuman ships a real-time animated desktop mascot with full voice capabilities:

  • STT in: dictate tasks and questions via microphone
  • ElevenLabs TTS out: the mascot speaks responses with lip-sync animation
  • Google Meet integration: the mascot joins a Google Meet call as a real participant — it listens to all speakers, takes notes directly into the Memory Tree, speaks back into the call when it has something to contribute, and pipes its animated face as the camera feed

Automatic Model Routing

A built-in router dispatches each task to the optimal LLM based on declared hints:

Routing Hint Target
hint:reasoning Frontier reasoning model
hint:fast Low-cost, low-latency model
vision tasks Vision-capable model

The default managed experience proxies all model calls through the OpenHuman backend under a single subscription — users do not manage separate API keys for Anthropic, OpenAI, Google, etc. Self-hosted API key configuration is also supported for 200+ models.

TokenJuice Compression

Every tool call result, scraped page, email body, and search payload passes through TokenJuice before reaching any LLM:

  • HTML-to-Markdown conversion
  • URL shortening and deduplication
  • Output deduplication across tool calls
  • Claimed reduction in cost and latency of up to 80%

Architecture

Component Details
Primary language Rust 62.1%, TypeScript 33.9%, JavaScript/Shell/CSS/HTML (remainder)
Frontend TypeScript + React 19
Desktop framework Tauri (Rust-based cross-platform desktop)
Memory backend SQLite on-device; Obsidian-compatible Markdown vault
Integration layer Composio connector (OAuth management + data ingestion)
Model access OpenHuman managed backend (default) or self-hosted API keys
Token compression TokenJuice (in-process, before LLM calls)
OS support macOS, Windows, Linux
License GPL-3.0

Installation

Platform Method
macOS / Linux (quick) curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh \| bash
macOS (recommended) DMG download from tinyhumans.ai/openhuman — verify developer signature before installing
Windows PowerShell: irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 \| iex

No terminal or config file required to get started — GUI onboarding flow handles first-run setup.

Local development requirements: Node.js 24+, pnpm 10.10.0, Rust 1.93.0, CMake, Ninja, ripgrep.

Privacy and Data Model

OpenHuman follows a local-first architecture: all memory, vault data, and SQLite state live on the user's machine. A local LLM handles low-level summarization and tooling to keep sensitive data off cloud endpoints.

The managed experience (default) does route some traffic through TinyHumans servers: - Model call proxying and routing - OAuth flow management via Composio - Web search proxying

Users who want fully air-gapped operation can configure self-hosted models and opt out of managed services, though integration OAuth flows then require self-managed Composio setup.

Security Considerations

OpenHuman inherits the same toxic flow trifecta risk as OpenClaw and Hermes: private data access + untrusted content handling + external action capability. Specific concerns:

Risk Description
Broad OAuth surface Simultaneous OAuth access to email, code repositories, calendar, chat history, and payment tooling — a misconfigured or compromised token can cascade across all connected services
Piped install risk The curl \| bash install pattern grants execution privileges to remotely hosted code before inspection. DMG download + developer signature verification is safer
Managed backend trust Default model routing and OAuth flows depend on TinyHumans servers — supply chain trust applies to the managed layer, not just the OSS client
Prompt injection 118+ auto-fetched data sources are potential injection vectors; adversarial content in emails, calendar events, or Slack messages could influence agent behavior

Mitigations

  • Use the DMG/installer download rather than curl | bash
  • Review OAuth scopes carefully; grant only integrations you actively need
  • Inspect the Obsidian vault periodically to audit what the agent has retained
  • Run on a dedicated machine or user account if handling sensitive credentials
  • For high-security environments, disable managed backend services and configure fully local models

Comparison with OpenClaw and Hermes Agent

Dimension OpenHuman OpenClaw Hermes Agent
Primary use case Personal AI with persistent life-memory Hyper-personal assistant Hyper-personal assistant with learning loop
Memory capacity 1B tokens, NeoCortex hierarchical tree 3-tier Markdown (MEMORY.md, daily, DREAMS.md) FTS5 + Honcho user modeling
Self-improvement ❌ ❌ ✅ Automatic skill creation + refinement
Integrations 118+ OAuth (Composio) 23+ messaging platforms MCP servers
Messaging gateways ❌ (no chat app bridge) ✅ WhatsApp, Telegram, Slack, iMessage, 20+ more ✅ Telegram, Discord, Slack, WhatsApp, Signal, Email
Voice ✅ STT + ElevenLabs TTS + mascot lip-sync ✅ ElevenLabs; wake word (macOS/iOS); continuous (Android) ✅ Voice memo transcription
Video meeting ✅ Google Meet as live participant ❌ ❌
Model flexibility 200+ via managed router or own keys Any provider + Ollama local Any provider via OpenRouter
Token compression ✅ TokenJuice (up to 80% reduction) ❌ ❌
Desktop app ✅ Native (Tauri) ❌ (CLI + terminal UI) ❌ (CLI + terminal UI)
MCP support ❌ ❌ ✅
Language Rust + TypeScript TypeScript / Node.js Python
Open source ✅ GPL-3.0 ✅ MIT ✅
GitHub stars (May 2026) ~27k (2 weeks post-launch) 373k ~153k
Security profile Toxic flow trifecta; managed backend dependency Toxic flow trifecta; active supply chain incidents Toxic flow trifecta (inherited)

Suitable For

  • Individuals wanting a persistent AI that knows their full personal and professional context across email, calendar, code, and documents
  • Teams evaluating memory-centric personal agent architectures with a 1B-token horizon
  • Developers who want a desktop-native agent with a polished GUI rather than a terminal-first experience
  • Users interested in voice-first or video meeting AI integration without a separate service layer
  • Builders who want an auditable, locally stored memory vault (Obsidian-compatible)

Limitations

  • No self-improvement or skill learning loop — the agent's behavior does not autonomously improve from past task outcomes (cf. Hermes Agent)
  • No native messaging gateway — cannot be reached via WhatsApp, Telegram, Slack, etc. (cf. OpenClaw, Hermes)
  • No native MCP support
  • Default experience depends on the TinyHumans managed backend for model routing and OAuth — partial cloud trust required
  • GPL-3.0 license is more restrictive than MIT (OpenClaw) for commercial embedding
  • Very new project (May 2026 beta): production track record and long-term ecosystem maturity are unproven
  • curl | bash install method is a security risk; DMG download is recommended

See Also

References