LobeHub is an open-source AI agent platform that has grown well beyond its origins as LobeChat. Its current focus is persistent AI coworkers: agents that can use different models, access tools and knowledge, remember information, collaborate with other agents, and keep working through projects or scheduled tasks. LobeHub describes the underlying system as an Agent Harness, which handles prompts, tools, planning, files, lifecycle behavior, and sub-agent coordination around the chosen AI model.

That makes LobeHub more ambitious than a multi-model chat interface. The main value is organizing AI into repeatable units of work rather than starting from an empty conversation every time.

Agents are the core building block. LobeHub lets you define a role, behavior, knowledge, and capabilities for each one. Its current platform can also generate an initial agent configuration from a short description, which lowers the setup burden for users who know what they want an agent to do but don't want to write every instruction manually.

Agents can be extended with Skills, reusable instruction packages that give them more specialized behavior. LobeHub also maintains large community marketplaces for agents, Skills, and MCP servers, so users can start from existing configurations instead of building everything themselves.

The marketplace approach is useful, but quality will naturally vary. Community agents and skills should be reviewed before being trusted with important workflows.

One of LobeHub's strongest ideas is Agent Groups. Instead of routing an entire task through one assistant, the platform can assemble several agents with different responsibilities.

The current system supports automatic team formation, parallel collaboration, multiple simultaneous tasks, and iterative improvement between agents.

For example, a research project could use one agent to collect material, another to analyze it, and another to prepare the final document. This is more useful than forcing a single agent to act as researcher, critic, writer, and editor at once.

Multi-agent work isn't automatically better, though. More agents introduce more coordination and more opportunities for one poor result to affect later steps.

LobeHub is also moving beyond chat into longer-running work.

Pages let several agents contribute within shared context. Projects organize related conversations and work. Schedules let agents run jobs automatically, while LobeHub's newer task system is designed around assigning work to agents, tracking progress, reviewing results, and handling recurring tasks. This makes the platform more relevant for recurring research, reporting, monitoring, writing, and operational workflows than a chatbot that only responds when opened.

Persistent memory is another important part of LobeHub's direction. The platform describes Personal Memory, continual learning, adaptive behavior, and White-Box Memory, meaning stored information is structured and editable rather than completely hidden from the user. That is a sensible approach for long-term agents. Memory becomes more useful when you can correct an outdated preference or remove something the agent should no longer rely on.

LobeHub remains model-flexible. Its platform is designed around multi-model orchestration and supports bringing different AI providers into the same environment. MCP extends agents further by connecting them to external tools and APIs.

Developers also get an official LobeHub CLI. It can create and run agents, manage files and knowledge bases, configure model providers, and generate text, images, video, speech, and transcriptions from the terminal. LobeHub can even be exposed as a callable tool inside other agent systems.

Desktop apps are available for macOS, Windows, and Linux, with mobile versions for iOS and Android.

LobeHub's biggest advantage is also its main drawback: there is a lot to configure. Agents, models, Skills, MCP servers, memory, knowledge, projects, schedules, and multi-agent groups give experienced users control, but they can feel excessive if all you need is a straightforward AI chat.

Persistent agents also require supervision. Incorrect memories, weak community Skills, excessive tool permissions, or poorly divided agent roles can produce problems that become harder to spot as workflows grow.