OpenHuman
OpenHuman is an open-source agent harness built around a Rust core. TinyHumans provides desktop, browser and terminal interfaces, plus a Rust library for embedding the runtime in another process. The repository labels the project early beta and warns users to expect rough edges.
Choose the engines behind the interface
The README describes configurable inference, embedding, memory and search engines. Model options include a managed TinyHumans route, local systems such as Ollama or LM Studio, and bring-your-own-key providers. A desktop installation therefore does not imply local inference: the chosen engine determines where model requests go.
Memory Trees on TinyCortex, mirrored as an Obsidian vault, are described as the default local memory setup. Settings can switch to hosted or separately configured remote memory services. Check inference, embeddings and search as well as memory storage before calling an installation local-only. Connected accounts can supply sensitive data even when the memory files live on your machine.
Managed services use TinyHumans credentials. Other providers require their own configuration and may bill separately. Provider compatibility does not establish an affiliation or make access free.
Review an automation before saving it
OpenHuman describes workflows as saved, typed graphs built on tinyflows. The graph can include agent calls, HTTP requests, code, conditions, loops and approvals, and can trigger on a schedule, an app event or a manual run. You review the proposed graph on a canvas before saving it.
That review step is valuable for identifying which actions can modify files or contact external services. Check input assumptions, permissions and approval points with a harmless test before scheduling the graph. The README describes consequential browser actions such as purchases or sends returning a confirmation requirement; that description is not a complete security audit of all tools.
Run journals and per-call accounting can help inspect what happened. Preserve enough evidence to distinguish a model's proposed action from a successful external operation.
Match the runtime to the deployment
The desktop app targets Windows, macOS and Linux. Developers can also embed the core with separate agent providers, working directories, access tiers and sandboxes. Compile-time feature gates control which capabilities ship in a custom build.
The repository publishes memory-footprint and startup measurements, but those are project benchmarks for specified harness workloads. They do not establish end-to-end inference speed or the resource cost of your models. The README also distinguishes a direction toward thousands of agents from a scale it has already measured.
OpenHuman uses GPL-3.0. Review that license before embedding or distributing a product, and begin with a narrow, read-only task. Early-beta software, connected services and executable agent tools deserve a permissions check before broad access to personal accounts or production systems.