openhuman

[!NOTE] 🎉 Within one week of launch, OpenHuman became the number one trending repository on GitHub for nine days in a row.

Early beta. OpenHuman is under active development, so expect rough edges.

---

Install

The easiest way is to download the desktop app from tinyhumans.ai/openhuman or the latest release. There is a .dmg for macOS, an .msi or .exe for Windows, and a .deb or .AppImage for Linux.

Prefer the terminal? The install script picks the right package for your system, checks its checksum and installs it:

# macOS and Linux
curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.ps1 | iex

To preview what the macOS and Linux script would do, end the command with bash -s -- --dry-run. Other options and troubleshooting are in INSTALL.md.

---

Why OpenHuman?

Most agent harnesses run one heavy process per agent and resend a big prompt on every call. OpenHuman does the same work with far less. It is the only feature-rich open-source harness built for large fleets of agents: 500 of them fit on a $10 server.

---

Major innovations

Most agent harnesses are a simple loop: send everything to the model, wait, repeat. That works for one agent, but it gets slow and expensive quickly, and it falls apart when you run hundreds.

OpenHuman rethinks the parts that cost the most: how much text the AI has to read, how it finds the right tool, how features load, and how much machine the whole thing needs. The six ideas below are where the speed and savings above come from. Each card links to the docs if you want the details.

---

Benchmarks

We ran seven harnesses on the same SWE-bench tasks, with the same model, API key and container, and metered every call. The setup and results are public in openhuman-benchmarks, so anyone can rerun them. The latest run is swe-x86-1, with ten tasks.

OpenHuman finished its tasks in less than half the median time, using 2.6x fewer tokens and an eighth of the memory and CPU. It made 114 model calls where the others made 134 to 212, and its whole run cost $0.05 where the others cost $0.08 to $0.35.

It wins by doing less work per step. The core is compiled Rust in a single process, not Node or Python, so it uses little memory and almost no CPU between model calls. It sends the smallest prompt of the seven, 4.6k tokens with its 20 tools included, so each call is cheaper and comes back faster (1.65 s median, also the fastest). It also needs fewer calls to reach an answer, which is where most of the savings come from.

Per solved task	OpenHuman (vs. median)	Median of the other six	Best of the other six
Wall time (p50)	19.8 s (2.3x faster)	46.5 s	28.4 s (OpenCode)
Tokens	167k (2.6x fewer)	435k	370k (Codex)
Cost	$0.0077 (-36%)	$0.012	$0.0077 (OpenCode, a tie)
Peak memory	68 MB (7.7x less)	523 MB	123 MB (Codex)
CPU time	1.3 s (8x less)	10.4 s	2.9 s (DeepSeek Harness)
System prompt	4.6k tokens (-40%)	7.7k	6.2k (DeepSeek Harness)

---

For users

If you use Claude Code, Codex, OpenClaw or Hermes, you already know the ideas: an agent loop with tools, MCP, skills, BYOK models and memory. OpenHuman has all of them, in a desktop app, a terminal app or a headless server.

Capability	What OpenHuman ships
Configurable memory	Built-in memory over your files, repos, feeds and apps, recalled before every turn with citations, on the memory engine you choose
Voice agents	A live voice agent you can interrupt mid-sentence, plus dictation and spoken replies
Computer and browser control	Drives a real Chrome browser and your desktop apps, asking before anything it cannot undo
Search	Search engines and search agents: Exa and Gemini included, Brave, Tavily, Parallel and more with your own key, grounded answers with citations and Deep Research
Tools	Native tools: shell and coder, scraper, documents, image and video generation, cron
MCP and skills	MCP servers and skill bundles
OAuth integrations	119 apps through Composio, with triggers that start the agent when something happens
Models	Local models (Ollama, LM Studio, MLX) and BYOK for 26 providers, with automatic model routing
Channels	14 messaging channels as agent front ends: Telegram, Discord, iMessage, email and more
Workflows	Durable workflow graphs: cron, event or manual triggers, approval steps, resume after a pause
Safety	Approval gate, sandboxed execution (OS jail or Docker), Privacy Mode for local-only runs, secrets in the OS keyring
Usage tracking	Per-call cost and token usage, plus replayable run journals

Start with Getting started or the guides.

---

For developers

OpenHuman is a library-first harness. Add it to your Rust project and call an agent like any other function, with no sidecar or daemon to run. One runtime can hold hundreds of agents, each with its own model, tools, memory and sandbox. 500 of them fit on a $10 VPS, so a product with one agent per customer can start without a cluster.

use openhuman_embed::{Access, Harness, Provider, Workspace};

let agent = Harness::builder()
    .provider(Provider::openai_compatible("https://api.openai.com/v1", "sk-...").model("gpt-5"))
    .workspace(Workspace::Ephemeral)
    .access(Access::readonly())
    .build()
    .await?;

let reply = agent.run("Summarize what you can see in this directory.").await?;
println!("{}", reply.reply);

Next: the Rust quickstart, the embedding guide and the developer docs.

---

How it compares?

Here is OpenHuman next to the harnesses most people already use. The top four rows come from our public benchmark: ten SWE-bench coding tasks, with the same model and container for every harness. Each number is per solved task. The rest compare features.

In short, the others are great for one person working with one agent. OpenHuman does that too, and it is also the open-source option you can embed as a library and scale to a fleet. Products change quickly, so check each project before you decide.

	Claude Code	Codex	OpenClaw	Hermes Agent	OpenHuman
Median task time	37.8 s	36.1 s	62 s	58.5 s	🚀 19.8 s
Avg tokens per task	428k	370k	442k	482k	🚀 167k
Peak memory per task	231 MB	123 MB	1.57 GB	816 MB	🚀 68 MB
Avg cost per task	$0.04	$0.02	$0.01	$0.0082	🚀 $0.0077
Open source	🚫 Proprietary	✅ Apache-2.0	✅ MIT	✅ MIT	✅ GPL-3.0
Agent fleets	⚠️ Process per agent	⚠️ Process per agent	⚠️ Process per agent	⚠️ Process per agent	🚀 500 agents on a $10 VPS
Embeddable library	⚠️ SDK over a CLI	⚠️ SDK over a CLI	🚫 None	⚠️ Python package	🚀 Typed Rust API
Memory	⚠️ Memory files	⚠️ Memory files	⚠️ Plugin-reliant	✅ Self-learning	🚀 Recalled every turn, with citations
Integrations	✅ MCP	✅ MCP	⚠️ BYO	⚠️ BYO	🚀 119 OAuth apps, MCP, skills
Messaging channels	🚫 None	🚫 None	✅ Many	✅ Several	✅ 14, including email
Browser and desktop	⚠️ Browser via MCP	⚠️ Browser via MCP	✅ Browser	✅ Browser	✅ Browser and desktop apps
Workflows	🚫 None	🚫 None	⚠️ Scripts	⚠️ Scripts	🚀 Visual, agent-drafted
Model choice	⚠️ Anthropic models	⚠️ OpenAI-first	✅ Any	✅ Any	✅ Any, with built-in routing

---

Contributing

Read CONTRIBUTING.md, or let an AI coding agent guide you with this prompt.

1. Install Git, Node.js 24+, pnpm 10.10.0, Rust 1.99.1 (with rustfmt and clippy), CMake, Ninja, ripgrep, and your platform's desktop build prerequisites.
2. Fork and clone the repo. Run git submodule update --init --recursive, then pnpm install.
3. Run pnpm dev for UI work or pnpm dev:app for the desktop app. Before you open a PR, run pnpm typecheck, pnpm format:check and cargo check --manifest-path Cargo.toml.

More in Getting set up, AGENTS.md and the crates overview. Many parts of OpenHuman live in their own repos under vendor/, and they welcome contributions too.

Contributors get free merch and special access on Discord.

Star history

Contributors