LobeHub

LobeHub is a Next.js application backed by PostgreSQL, with optional Redis and S3-compatible storage; it also ships desktop and mobile clients and an npm CLI (lh). The chat surface, the model endpoint and the place where tools execute can be different machines: the docs separate the interface from an execution device.

For multi-agent work there are three layers. Agent Groups place specialist agents in one group chat in sequential, parallel, iterative or debate mode, and an included Orchestrator understands the goal, assigns work and summarises. Tasks hand a piece of work to one agent to run in the background, with subtasks, dependencies, comments, acceptance checks and cron-style schedules. External agents (coding CLIs, OpenClaw, Hermes) are attached as separate LobeHub agents, and each keeps its own session, project instructions and permissions.

State lives in the database as topics, tasks, pages and agent documents. The README describes per-agent memory that the user can view and edit. For external agents the docs say session recovery depends on the external runtime, and OpenClaw and Hermes keep their own workspace and memory outside the LobeHub topic.

Each connector tool can be set to run directly, ask for approval, or stay off, and the docs say these settings do not replace an external agent's own permission system. Tasks can pause in a Pending review status, carry acceptance criteria that can hold back delivery and start a repair round, and take comments for follow-ups. Chat shows tool calls, diffs and questions or approval requests from external agents where the integration supports them. The docs advise starting with read-only tasks and reviewing file changes before the next step.

bash <(curl -fsSL https://lobe.li/setup.sh)

# Self-host with Docker (README steps)
mkdir lobehub-db && cd lobehub-db
bash <(curl -fsSL https://lobe.li/setup.sh)   # writes the compose files and env for the LobeHub stack
docker compose up -d

# Optional: use another computer as an execution device for external agents
npm i -g @lobehub/cli
lh connect

# In the app: Create Agent -> Connect External Agents, choose the device, Rescan, Connect.

• The README's shell snippets use fish syntax ($ mkdir ...); the setup line is bash process substitution and needs bash or zsh.
• An LLM API key is required to chat; the README's one-click deploy asks for OPENAI_API_KEY.
• External agents must be installed and signed in on the execution device itself; the connection wizard scans that device and does not create a cloud sandbox.
• A working directory is not a security boundary, and browser-paired devices offer fewer tools than the desktop app.
• Some features are labelled Alpha or Labs and can be missing or renamed between versions (see the feature status page).

The source is public on GitHub, but GitHub reports the license as unknown because the LICENSE file is the LobeHub Community License: Apache 2.0 with added conditions. Commercial use as a service without modifying the source is allowed; developing and distributing a derivative work needs a commercial license from LobeHub.

Yes as a client. The Skill Store has an MCP tab for community integrations and lets you add a custom MCP endpoint. Local MCP connectors run through the desktop app.

It attaches them rather than reimplementing them. The Connect External Agents wizard scans a device for installed CLIs, including Claude Code, Codex and OpenCode, and also OpenClaw and Hermes, and each becomes a separate LobeHub agent that keeps its own session and permissions.

Through Agent Groups, where a built-in Orchestrator coordinates specialist agents in sequential, parallel, iterative or debate mode, and through Tasks that delegate background work to a single agent.

Self-hosting the app has no license fee within the license terms, and LobeHub Cloud lists a free plan plus paid plans on its pricing page. Model API usage is billed by your provider when you bring your own key.