The agent marketplace: 500+ pre-built agents
LobeHub's strongest feature is the agent marketplace. It has pre-built agents for: code review on GitHub PRs, daily Hacker News summarization, customer support auto-reply, meeting note generation, and 500+ more. Each agent is a prompt template with tool configurations, model settings, and scheduling rules. I deployed the 'GitHub Issue Triage' agent in 2 clicks: select it from the marketplace, configure my GitHub token, set it to run every 2 hours. It started triaging real issues within 5 minutes. No code required. This marketplace experience is better than anything AutoGPT or CrewAI offer.
Scheduling agents: the promise vs the reality
I scheduled 3 agents: GitHub triage (every 2 hours), daily digest (8 AM), and repo health check (every 6 hours). In 21 days: 42 scheduled runs, 8 missed runs (19% failure rate). The failures were silent: no alert, no retry, the agent just did not fire. When the agents ran, they performed correctly 85% of the time. The other 15% produced output that was formatted wrong (JSON instead of markdown) or contained hallucinated GitHub issue numbers. For non-critical automation it is fine. For anything where a missed run costs money, do not rely on the built-in scheduler yet.
The chat interface: better than ChatGPT for power users
LobeHub's chat UI is genuinely better than ChatGPT's in several ways. Plugin system: install plugins for web search, image generation, code execution, and file reading directly into the chat interface. The web search plugin uses SearXNG (privacy-respecting metasearch) instead of Bing. Multiple model providers in one dropdown. Conversation forking: branch a conversation at any point to explore alternative responses. Prompt library: save and reuse prompt templates. The UI is built with React and is fast, responsive, and feature-rich. This alone makes LobeHub worth using even without the agent features.
Self-hosting vs cloud: the server requirements
Self-host with Docker: `docker run -d -p 3210:3210 lobehub/lobe-chat`. It needs a PostgreSQL database for agent state and conversation history, and an S3-compatible storage for file uploads. The Docker Compose file sets all of this up. I ran it on my HK server (2 vCPU, 4GB RAM) and it used 1.2GB RAM at idle, 2.5GB with 3 active agents. The cloud version (lobehub.com) offers a hosted experience with the same features. Self-hosting gives you data control. The cloud version gives you one-click setup. Both are valid.
What LobeHub gets right that others miss
The plugin architecture is the most underrated feature. Each agent can use plugins: the GitHub triage agent uses a GitHub API plugin with read/write permissions, the daily digest agent uses a web search plugin, the health check agent uses a shell execution plugin. These plugins are scoped: each agent only gets access to the plugins it needs. This is better than Claude Code or AutoGPT where the agent gets full shell access by default. LobeHub's plugin system is the right security model for autonomous agents.
