OpenHuman is an agent harness built to run large fleets of agents cheaply. It ships as a desktop app, a terminal app, and a headless server for people, and as a Rust library for developers, who call an agent like any other function with no sidecar or daemon. One runtime holds hundreds of agents, each with its own model, tools, memory, and sandbox, and more than 500 fit on a $10 VPS.
The core is compiled Rust in a single process that starts in a tenth of a second. Large tool results are compressed with Recursive Language Models (RLM) before the model reads them, or handed over as a searchable handle when very large. Jev, a small model, picks the right tool out of 1,215 (86.8% top-pick accuracy against 70.5% for keyword search) and also chooses on-screen buttons for browser and desktop control without screenshots. Features such as search, documents, and voice plug into one Rust bus and load only when needed, and built-in memory selects relevant context within a token budget before every turn and cites it.
On ten SWE-bench tasks, run on seven harnesses with the same model, API key, and container, the published per-task results were:
| Per solved task | OpenHuman | Median of the other six |
|---|---|---|
| Wall time (p50) | 19.8 s | 46.5 s |
| Tokens | 167k | 435k |
| Cost | $0.0077 | $0.012 |
| Peak memory | 68 MB | 523 MB |
| CPU time | 1.3 s | 10.4 s |
| System prompt | 4.6k tokens | 7.7k |
For users it covers the same ground as Claude Code, Codex, OpenClaw, and Hermes Agent: an agent loop with tools, MCP, skills, bring-your-own-key models, and memory, plus voice agents, 14 messaging channels, 119 OAuth apps through Composio with triggers, durable workflow graphs with approval steps, and per-call cost tracking. Local models run through Ollama, LM Studio, or MLX, with automatic model routing across 26 providers. It is in early beta.
Features
- Agent fleets: hundreds of agents in one runtime, each with its own model, tools, memory, and sandbox
- Rust library: a typed API that embeds the harness in a Rust project with no sidecar
- RLM token compression: large tool results are compressed or exposed as a searchable handle
- Jev tool search: a small model that selects tools and on-screen controls
- Built-in memory: memory over files, repos, feeds, and apps, recalled before every turn with citations and a swappable engine
- Browser and desktop control: drives a real Chrome browser and desktop apps, asking before irreversible actions
- Search: Exa and Gemini built in, more engines with your own key, and Deep Research with grounded citations
- Integrations: MCP servers, skill bundles, and 119 OAuth apps through Composio with event triggers
- Models: local models through Ollama, LM Studio, or MLX, bring your own key for 26 providers, and automatic routing
- Channels: 14 messaging front ends including Telegram, Discord, iMessage, and email
- Workflows: durable graphs with cron, event, or manual triggers, approval steps, and resume after a pause
- Safety: an approval gate, OS-jail or Docker sandboxing, a local-only Privacy Mode, and secrets in the OS keyring
