LobeHub

What it is

LobeHub is a workspace for running a team of AI agents. Its README calls it your Chief Agent Operator, one that “hires, schedules, reports on your entire AI team” so it keeps working while you are away. You make an agent by describing it to the Agent Builder, or add one the community shares. Agents join groups that work in turn, side by side, in rounds or in a debate. You give an agent a task, and a schedule can rerun it. Memory is kept as entries you can read and edit. Agents reach tools through connectors and MCP servers, and each agent can use its own model. LobeHub can also take in coding agents that run on your computer, such as Claude Code and Codex. This listing covers running LobeHub yourself, and the apps.

1. You describe an agent, or add one from the community
2. You hand it a task, or put it in a group
3. It works with its own model, tools and memory
4. A schedule reruns the task when you set one
5. You review the run record and the output

Who it suits

People who want several AI agents in one place, with their chats, files and memory kept together, and teams that want to host that place on their own server. A developer who already uses Claude Code, Codex or Cursor can connect them in the same workspace. Self-hosting means looking after a database, file storage and the app itself, so it asks more of you than the web app does.

Skip the self-hosted route if you do not want to run a database, file storage and a server. The web app at app.lobehub.com and the desktop app work without them.

Where to use it

LobeHub runs where you put it, and the docs name these places.

• Your own server with Docker Compose, which the docs recommend, or one Docker container with a database you provide
• Vercel, Zeabur, Sealos, RepoCloud, Dokploy and Alibaba Cloud
• The web app at app.lobehub.com, the macOS, Windows and Linux desktop apps, and the iOS and Android downloads
• The source code, if you want to change it
• Claude Code, Codex, Cursor and the other coding agents on its list, connected from a desktop or from a computer joined with the LobeHub CLI

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with a file for each route and one on connecting agents. The main route is Docker Compose on Linux or macOS. Windows runs it through WSL 2.

1. Make a folder and open it in a terminal.
2. Run the setup script. Press Enter to take the default, local mode, which only this computer can reach. Port mode and domain mode are for a network or a domain.
3. Read the report the script prints. It lists the addresses and the file store’s user name and password.
4. Start the stack and watch the log.
5. Open http://localhost:3210.

mkdir lobehub && cd lobehub

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

docker compose up -d

docker logs -f lobehub

The docs say the log shows a database migration line when the app has started. The script is a file fetched from lobe.li each time, so download and read it first if you prefer. Vercel and Dokploy need a database, storage and secrets you supply. The Zeabur and Sealos templates include PostgreSQL and storage and set up a Logto login, and RepoCloud is a one-click deploy. The steps for each are in the kit. A single container needs a PostgreSQL database with the pgvector and pg_search extensions (the Docker page uses paradedb/paradedb:latest-pg17, and the 2.0 notes recommend PostgreSQL 17 or higher) plus S3-compatible storage. It takes your settings from a file.

How to use it

Open LobeHub and start in Lobe AI, the built-in agent, which needs no agent setup. For work you repeat, make your own.

1. Click Create Agent in the left sidebar and describe the job to Agent Builder in the right panel. It writes the role, the system prompt and the skill setup, and you adjust them.
2. For a team, click Create Group. Start from a template or assemble your own team, and choose whether to include an Orchestrator and its model. The docs describe four ways a group works: sequential, parallel, iterative and debate.
3. For work that needs follow-up, open Tasks on an agent’s page, describe the job and click Create task.
4. To repeat it, open the task’s detail, go to Automation, then Scheduled. Set the frequency, time, timezone and run limit, enable automation, start scheduling with the control the page shows, then check the next run time.

The docs advise running a task once by hand before you turn its schedule on, so you can check the agent has the tools and files it needs.

Example

You self-host LobeHub and have a file of customer feedback. On an agent’s Tasks page you type this and attach the file with the attachment icon.

Read the attached customer feedback, group it by issue type, and identify the three most common problems. Suggest an improvement for each problem and present the results in a table.

That is the docs’ own sample task. You click Create task, and it appears in the list under a status such as Backlog or Pending review. You open it to see the run records and the output. If the answer is what you wanted, the next step in the docs is Automation, where you choose how often it runs.

For newcomers and experts

If you are new. An agent here is a saved role: a system prompt, a model and the tools it may use. A topic is one conversation with it. A task is closer to a ticket. You write what you want, assign an agent and let it work in the background, and the docs compare it to Linear or GitHub Issues with an agent as the assignee. Self-hosting means you run the app, its database and its file storage yourself. Start with the Docker Compose route in local mode, since it only answers on your own computer, and check the log lines the docs list before you open the page.

For experts. The app is Next.js and React on PostgreSQL through Drizzle, with tRPC for its API. The Docker image runs the database migration before it starts. The Compose stack adds a gateway for Gateway Mode, which runs agent work on the server instead of in the browser. It needs AGENT_GATEWAY_URL, GATEWAY_SERVICE_TOKEN and a pair of keys, JWKS_KEY and JWKS_PUBLIC_KEY. The browser opens a WebSocket to the gateway, so a proxy in front of it must allow upgrades. Port 8788 serves the device gateway, which lets a computer joined with lh connect run local tools.

The repository also holds an OpenAPI file and a TypeScript SDK, @lobehub/sdk. The docs say to check the version you run before you build on them, and not to treat the internal chat and tRPC routes as a stable public interface.

Good to know

LobeHub runs on your own server, and a few details are worth reading first.

• License: the LobeHub Community License, which GitHub lists as Other. The file says it is based on Apache License 2.0 with added conditions. It allows commercial use as a front end and back end without changing the source. It asks for a commercial license from the producer to develop and distribute a derivative work. Contributors agree the producer can change the terms, and that their code may be used commercially, including in its cloud edition. The text names the product LobeChat
• Secrets: KEY_VAULTS_SECRET encrypts the API keys users store. Once it is set, do not change it, or that data cannot be decrypted. Every deployment needs its own JWKS_KEY
• Updating: docker compose pull and docker compose up -d update the stack. An older Compose deployment needs the new gateway variables first, or the gateway container keeps restarting and agent runs fall back to the browser
• Vercel: the docs list a 10 second function limit on Hobby and no WebSocket support, and it needs an outside database and storage
• Pages that disagree: the Zeabur and Sealos pages set up Logto, while the Docker and Vercel pages use the built-in login. The README’s buttons for both hosts point to different templates from the ones the pages link
• Connected agents: a coding agent keeps its own permission rules, and LobeHub’s connector permissions do not replace them. The docs say a working directory is not a security sandbox
• Schedules: scheduled runs can be daily, weekly or hourly. Monthly and custom cron are not offered. If runs or costs get out of hand, the docs say to disable automation first
• Your data: prompts and files go to the model providers you set. The docs list optional analytics from Vercel, Google, PostHog and Umami, each turned on by its own variable, and say these work only on Vercel deployments, not Docker or Docker Compose. The same page lists Langfuse, which traces chat data to a Langfuse server once its keys are set (ENABLE_LANGFUSE defaults to 1)
• Removal: docker compose stop stops the stack. The docs give no uninstall page, and their database reset deletes the ./data folder