OpenHuman AI Agent Complete Tutorial: Build Your Private AI Assistant from Scratch

Created

OpenHuman AI Agent Complete Tutorial: Build Your Private AI Assistant from Scratch

🤖 OpenHuman is your Personal AI super intelligence. Private, Simple and extremely powerful.

💡 Key Takeaways

  1. OpenHuman is an open-source desktop-grade AI assistant framework with core logic written in Rust
  2. Built-in 118+ third-party integrations (Gmail, Notion, GitHub, Slack, etc.) with one-click OAuth
  3. Memory Tree technology lets AI understand your work context in minutes
  4. TokenJuice intelligent compression reduces token costs and latency by up to 80%
  5. Supports local AI via Ollama or LM Studio
  6. Single-subscription multi-model architecture with built-in intelligent model routing
  7. Licensed under GPL-3.0, with fully auditable source code

📖 Project Overview

What is OpenHuman?

OpenHuman is an open-source Personal AI super intelligence framework with core logic written in Rust, building cross-platform desktop applications through Tauri[^1]. Developed by the TinyHumans AI team, it aims to seamlessly integrate AI capabilities into daily life, enabling agents to understand your work context in minutes rather than requiring days or weeks of training like other agent frameworks[^1].

Key Statistics[^2]:

  • 🌟 21,096+ GitHub Stars (as of May 19, 2026)
  • 🍴 1,857+ Forks
  • 📅 Project Created: February 18, 2026
  • ⚖️ License: GPL-3.0
  • 🏷️ Current Status: Early Beta — actively under development

Core Features:

  • 🔒 Privacy — Runs locally, workflow data stays on your device with local encryption[^1]
  • 🚀 High Performance — Rust core with TokenJuice intelligent compression reducing token costs and latency by up to 80%[^1]
  • 📦 Ease of Use — UI-first design, ready to use in just a few clicks after installation, no terminal required[^1]
  • 🎯 Feature-rich — Supports multimodal tasks, voice interaction, and Google Meet meeting integration[^1]

[^1]: Source: OpenHuman GitHub README
[^2]: Source: GitHub API

📊 Key Features

1. UI-First Design

OpenHuman takes a UI-first approach, letting you go from installation to a working agent in just a few clicks — no config-first, no terminal[^1].

The agent features a desktop mascot: it talks, reacts, joins your Google Meet meetings, remembers you, and continues thinking in the background after you stop typing[^1].

Key Insight: OpenHuman is the first agent framework that understands you in minutes. Most agents start from zero — Hermes learns by observing your work; OpenClaw waits for plugins to feed context. Either way, you spend days or weeks[^1].

2. 118+ Third-Party Integrations

Connect to Gmail, Notion, GitHub, Slack, Stripe, Calendar, Drive, Linear, Jira, and more with one-click OAuth[^1]. Each connection is exposed as a typed tool for the agent, with the core proactively fetching the latest data into the Memory Tree every 20 minutes. No prompts needed, no polling loops[^1].

3. Memory Tree + Obsidian Wiki

A local-first knowledge base built from your data and activity[^1]. All connected content is normalized into ≤3k token Markdown blocks, scored and folded into a hierarchical summary tree stored in local SQLite[^1]. The same blocks are stored as .md files in an Obsidian-compatible vault that you can open, browse, and edit[^1].

Inspiration: This design draws from Karpathy’s obsidian-wiki workflow.

4. Built-in Toolset

Out of the box[^1]:

  • 🔍 Web Search
  • 🌐 Web Scraper
  • 💻 Full Coding Toolset (filesystem, git, lint, test, grep)
  • 🎤 Native Voice (STT input → ElevenLabs TTS output, mascot lip sync, Google Meet Agent)

Model routing sends each task to the right LLM (reasoning, fast, or vision) with a single subscription[^1]. There’s also an option for local AI via Ollama for local workloads[^1].

5. TokenJuice Intelligent Compression

Every tool call, scrape result, email body, and search payload runs through the token compression layer before touching any LLM[^1]. HTML converts to Markdown, long URLs shorten, verbose tool outputs get deduplicated and summarized via configurable rules… CJK, emoji, and other multi-byte text is preserved glyph-by-glyph — never stripped[^1]. Same information, fewer tokens, cutting costs and latency by up to 80%[^1].

6. Messaging Channels

Inbound/outbound interaction through the channels you already use, with workflow data staying on your device, locally encrypted, treated as your property[^1].

7. Optional agentmemory Backend

If you’ve self-hosted agentmemory in other coding agents, OpenHuman provides an optional Memory backend to proxy it — set memory.backend = "agentmemory" in config.toml, and the same persistent store serves both OpenHuman and Claude Code, Cursor, Codex, OpenCode[^1].

📦 System Requirements

Item Requirement
Operating System Linux / macOS / Windows
Programming Language Rust 1.93.0+
Memory At least 4GB RAM
Disk Space 2GB available
Dependencies Git, Node.js 24+, pnpm 10.10.0, CMake, Ninja, ripgrep

🚀 Quick Start

macOS or Linux x64:

1
curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash

Windows:

1
irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 | iex

Method 2: Download from Website

Visit tinyhumans.ai/openhuman to download the DMG or EXE installer.

💻 Development Build

Build from Source

1
2
3
4
5
6
7
8
9
10
11
12
# Clone the repository
git clone <repo-url>
cd openhuman

# Initialize submodules
git submodule update --init --recursive

# Install dependencies
pnpm install

# Run in development mode
pnpm dev

Build Commands

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# Build Web frontend (pure web mode)
pnpm build:web

# Universal build (Web + Desktop)
pnpm build

# Build desktop application
pnpm --filter openhuman-app build:app

# Type checking
pnpm typecheck

# Format checking
pnpm format:check

# Cargo check
cargo check -p openhuman --lib

🔧 Advanced Configuration

Model Routing Configuration

OpenHuman uses a single-subscription multi-model architecture with a built-in router that automatically dispatches tasks to the appropriate LLM. Configure the hint routing table via config.toml (note: the official README uses hint_reasoning format, not hint:reasoning prefix)[^1]:

1
2
3
4
5
[model_routing]
hint_reasoning = "anthropic/claude-sonnet-4"
hint_fast = "openrouter/google/gemini-2.0-flash"
hint_code = "anthropic/claude-sonnet-4"
hint_vision = "openrouter/google/gemini-2.0-flash"

Routing Table Reference:

Hint Key Typical Target Model Use Case
hint_reasoning Strong reasoning models (e.g., Claude Sonnet 4) Multi-step planning, math, heavy code tasks
hint_fast Fast & cheap models (e.g., Gemini 2.0 Flash) UI assistance, autocomplete, light classification
hint_code Code-tuned models Native coding tasks
hint_vision Vision models Screenshots, image attachments, OCR
hint_summarize Models good at compression Memory Tree summary building

📖 Official Docs: Automatic Model Routing

How It Works: When you use the hint:reasoning prefix in a chat message, the router looks up the config table and resolves to the corresponding model. All routing is proxied through OpenHuman’s single custom subscription backend, no need to manage multiple API Keys. You can also specify model names directly (without the hint: prefix) in a single call for precise control.

Local AI Configuration (Optional)

OpenHuman supports enabling local AI through Ollama or LM Studio for privacy-sensitive tasks:

1
2
3
# After installing Ollama, pull the default local models
ollama pull all-minilm:latest    # For Memory Tree embedding vector search
ollama pull gemma3:1b-it-qat     # For Memory Tree summary generation

Local AI Workloads (disabled by default, requires manual enablement):

Workload Type Default Model Purpose
Memory Embeddings all-minilm:latest Memory Tree vector search
Summary Generation gemma3:1b-it-qat Memory Tree source/topic/global summaries
Heartbeat Loop Configurable Background reasoning loop

Note: Chat, reasoning, and vision tasks still use cloud front-end models. This is intentional — local AI focuses on low-latency, privacy-sensitive repetitive work.

❓ Frequently Asked Questions

Q: What’s the difference between OpenHuman and Cursor, Windsurf, or Claude Code?

A: OpenHuman is a complete desktop-grade AI Agent framework, while Cursor, Windsurf, and Claude Code are essentially code editors or coding assistants. OpenHuman’s core advantages are: ① 118+ third-party integrations (Gmail, Notion, Slack, etc.), ② Memory Tree technology that lets AI understand your work context in minutes, ③ Desktop mascot for human-like interaction, ④ Google Meet meeting integration support. It’s more of a general-purpose AI assistant rather than being limited to coding scenarios.

Q: How do I run OpenHuman locally?

A: Download and extract to run — no additional software installation required. macOS and Linux users can deploy quickly with a one-click terminal install script, while Windows users can install via PowerShell. OpenHuman’s UI-first design ensures you’re ready to go in just a few clicks, with zero terminal operations needed. All workflow data stays on your device with local encryption for data privacy.

Q: Can OpenHuman connect to my data? How?

A: Yes, OpenHuman supports 118+ third-party app integrations including Gmail, Notion, GitHub, Slack, Stripe, Google Calendar, Google Drive, Linear, Jira, and more[^1]. Each connection is completed via one-click OAuth authentication and automatically exposed as a typed tool for the agent. The core proactively fetches the latest data into the Memory Tree every 20 minutes — no manual prompting or polling loops required.

Q: Does OpenHuman require an internet connection? Can it work fully offline?

A: OpenHuman needs an internet connection to access APIs and search functionality, but its core data (Memory Tree, workflow data) is stored entirely on your local device. If you’ve configured local AI models via Ollama or LM Studio, some reasoning tasks can run in an offline state. However, model routing and third-party integration features still require a network connection.

Q: What AI models does OpenHuman support? Can I customize models?

A: OpenHuman has a built-in model routing system that proxies multiple LLMs through a single custom subscription backend, including Anthropic Claude, Google Gemini, and more[^1]. You can customize the target model for each hint route via config.toml, or specify model names directly for precise control. Additionally, local AI models via Ollama or LM Studio are supported for privacy-sensitive tasks.

Q: What is a Memory Tree and how does it work?

A: The Memory Tree is OpenHuman’s local-first knowledge base, automatically built from your data and activity[^1]. All connected content is normalized into Markdown blocks of up to 3,000 tokens, scored and folded into a hierarchical summary tree stored in local SQLite. The same blocks are stored as .md files in an Obsidian-compatible vault that you can open, browse, and edit directly. This design draws inspiration from Karpathy’s obsidian-wiki workflow.

Q: How secure is OpenHuman? Is my data safe?

A: OpenHuman employs a multi-layer security design: ① All workflow data stays on your device with local encryption[^1], ② Memory Tree is stored in local SQLite without cloud dependency, ③ Third-party integrations authenticate via OAuth, so data is never directly exposed to AI models, ④ Licensed under GPL-3.0 with fully auditable source code.

Q: What development stage is OpenHuman currently in?

A: OpenHuman is currently in the Early Beta stage, actively under development[^1]. This means core features are functional, but some features may still be in progress. It’s recommended to back up your data when using it in production environments.

[^1]: Source: OpenHuman GitHub README

📊 OpenHuman vs Other AI Agents Comparison

Feature OpenHuman Cursor Windsurf Claude Code
Core Positioning Desktop AI Assistant Code Editor IDE Enhanced Coding CLI Tool
Third-Party Integrations 118+ (Gmail, Notion, Slack…) Limited (GitHub-focused) Limited Limited
Memory System Memory Tree (local SQLite + Obsidian) None None Yes (session-level)
Desktop Mascot ✅ Human-like interaction ❌ ❌ ❌
Voice Interaction ✅ STT + ElevenLabs TTS ❌ ❌ ❌
Meeting Integration ✅ Google Meet ❌ ❌ ❌
Local AI ✅ Ollama / LM Studio ❌ ❌ ❌
Cross-Platform ✅ Linux / macOS / Windows ✅ ✅ ✅
Open Source ✅ GPL-3.0 ❌ Proprietary ❌ Proprietary ❌ Proprietary
Model Routing ✅ Single-subscription multi-model ❌ ❌ ❌

💡 Summary: If you need a general-purpose AI assistant (not limited to coding), OpenHuman is currently the most comprehensive open-source option. If you focus on coding, Cursor or Windsurf might be a better fit.

Type Link
📖 Official Docs https://tinyhumans.gitbook.io/openhuman/
🐦 Twitter/X https://x.com/intent/follow?screen_name=tinyhumansai
💬 Discord https://discord.tinyhumans.ai/
📦 GitHub https://github.com/tinyhumansai/openhuman
📄 License GPL-3.0

👤 Who Is This Tutorial For?

This tutorial is suitable for:

  • 🔧 Developers: Who want to build their own AI assistant or integrate OpenHuman into their workflow
  • 🧪 AI Researchers: Interested in AI Agent architecture and memory systems
  • 🏢 Team Leads: Evaluating whether OpenHuman fits their team’s needs
  • 🙋 General Users: Who want a personal AI assistant that understands their work context

⚠️ Prerequisites: This tutorial assumes basic command-line proficiency and Git experience. If you’re new to programming, consider learning basic frontend development first.

🏁 Conclusion

OpenHuman is a powerful open-source AI Agent framework suitable for:

  • ✅ Developers building their own AI assistants
  • ✅ Researchers conducting AI experiments
  • ✅ General Users automating daily tasks

Start your AI Agent journey! 🚀

💬 Engage: If you’ve tried OpenHuman, feel free to share your experience on GitHub Issues or Discord!

📝 Article Information

Item Details
📅 First Published 2026-05-21
🔍 Information Verified 2026-05-21 (based on OpenHuman GitHub README and official docs)
📊 GitHub Stars 21,096+ (as of 2026-05-19)
📖 References OpenHuman GitHub, Official Docs
📝 Author Xiaoxin Blog Team
🏷️ License GPL-3.0