Subscribe on YouTube

How to Self-Host LobeHub

Tested with LobeHub 2.3.3 (Docker image `lobehub/lobehub:latest` on 2026-11-16, Agent Gateway 0.5.0). To get exactly that version, use the image tag `lobehub/lobehub:2.3.3` instead of `latest`.

LobeHub tutorial video6:38Watch the LobeHub tutorial on YouTube
Prefer watching? Subscribe for a new self-hosting tutorial every week.Subscribe on YouTube

LobeHub (formerly LobeChat) is a self-hostable chat and agent workspace for AI models, used from your browser. In this guide you self-host it with the official Docker Compose setup script, connect free local models through Ollama, then try chat, switching models, agents and topics.

Tested with LobeHub 2.3.3 (Docker image lobehub/lobehub:latest on 2026-11-16, Agent Gateway 0.5.0). To get exactly that version, use the image tag lobehub/lobehub:2.3.3 instead of latest.

Watch the video: https://youtu.be/XsjdkBC9h3A

Requirements

  • Linux or macOS, or Windows through WSL 2
  • Docker and Docker Compose
  • Minimum (light usage, from the docs): 2 CPU cores, 4 GB RAM, 20 GB disk
  • Free ports: 3210, 9000, 9001, 8787, 8788 (plus 11434 for Ollama in this guide)
  • For the local models: a graphics card helps (the Ollama command below uses an NVIDIA GPU via --gpus=all), and about 7 GB of disk for the two models
  • No cloud account and no paid API key

Step 1: Run the official setup script

Make a folder and run the official setup script with -l en for English. It downloads docker-compose.yml, searxng-settings.yml, bucket.config.json and .env.

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

Answer the prompts:

  • Deployment mode: press Enter (default, local mode)
  • "Do you want to regenerate security secrets? (y/n) [y]": press Enter
  • "Do you want to initialize the database? (y/n) [y]": press Enter (pulls the images, starts Postgres briefly, stops it)

Expected output (end):

text
Security secret generation results are as follows:
LobeHub:
  - URL: http://localhost:3210

The report also lists RustFS (port 9000) and the Agent Gateway (port 8787). The script writes freshly generated secrets into .env; keep that file private.

Step 2: Start the stack

bash
docker compose up -d
docker logs -f lobehub

Wait until the log shows the database migration and "Ready":

text
[Database] Start to migration...
✅ database migration pass.
...
- Local:        http://localhost:3210
✓ Ready in ...

Press Ctrl+C to stop following the logs (the app keeps running), then:

bash
docker compose ps

Every service (lobehub, lobe-gateway, lobe-postgres, lobe-redis, lobe-rustfs, lobe-searxng) is listed as Up. lobe-rustfs-init exits after its one-time job, which is normal.

Step 3: Create your local login

Open http://localhost:3210. This is your own instance, so the account lives only here.

  1. Click Sign up now.
  2. Enter an email address and a password (twice), click Sign Up, and agree to the terms.
  3. A short welcome follows (usage data opt-out, language, your name, interests, suggested agents). Choose what you like or skip.
  4. You land in the workspace with Lobe AI on the left. It has no model yet.

Step 4: Start Ollama with two free models

Start Ollama with the Docker command from the LobeHub Ollama docs. It uses your GPU, keeps models in the ollama volume, opens port 11434 and allows cross-origin requests:

bash
docker run -d --gpus=all -v ollama:/root/.ollama -e OLLAMA_ORIGINS="*" -p 11434:11434 --name ollama ollama/ollama
docker exec -it ollama ollama pull llama3.2
docker exec -it ollama ollama pull qwen3

Expected output (each pull ends with):

text
verifying sha256 digest
writing manifest
removing unused layers
success

llama3.2 is about 2 GB, qwen3 about 5.2 GB. No NVIDIA GPU? Drop --gpus=all (models then run on the CPU, slower).

So that the LobeHub container can reach Ollama by name (http://ollama:11434, used in the next step), attach Ollama to the stack's network. The network is <folder>_lobe-network, so with a folder called lobehub:

bash
docker network connect lobehub_lobe-network ollama

Step 5: Connect Ollama to LobeHub

  1. Click your name, open Settings, then Provider, and pick Ollama (it is already switched on).
  2. Set the address to http://ollama:11434.
  3. Click Fetch models: llama3.2 and qwen3 appear. Switch both on.
  4. Choose llama3.2 next to the Check button and click Check: it passes. (The default check model is one you have not pulled, so pick llama3.2 first.)

Step 6: Feature one - chat

On the home page, click the model name under the input box, open Model, and choose llama3.2. Type:

text
Explain what a Docker volume is, in two short sentences.

Press Enter. The answer streams in, generated on your machine, and the chat is saved as a topic. Hover over the answer to copy it, edit it, or ask again.

Step 7: Feature two - switch models

Click the model name again and pick qwen3. Ask:

text
Write a three-line poem about a cup of tea.

Qwen3 thinks first; you can expand its reasoning, then it answers. You can switch models any time inside the same chat.

Step 8: Feature three - agents

  • Click Community in the sidebar and search (for example "Spanish"). Shared agents are sorted by category. Opening and adding community agents needs a LobeHub account sign-in.
  • Build your own instead: click Add Agent, give it a name and create it. On its profile, pick its model (for example qwen3) and write its core instructions below.

Step 9: Feature four - topics

Every conversation is saved as a topic of its agent.

  • Open the three dots next to the chat title and choose Add to Home Favorites.
  • Open the menu again, choose Rename, and call it Docker notes.
  • Click the share button to export the chat as an image, Markdown, PDF or JSON.

Everything is stored in your own database.

Troubleshooting and tips

  • The setup script is for the first install only. After that, change settings in .env (see the "Custom deployment" part of the docs).
  • Port already in use: change LOBE_PORT in .env.
  • Database errors at the first start: the docs reset it like this (this deletes your chats and settings): ``bash docker compose down sudo rm -rf ./data docker compose up -d
  • Ollama will not connect: check that Ollama was started with OLLAMA_ORIGINS="*", that the address in the provider settings is right, and (for http://ollama:11434) that Ollama is attached to lobehub_lobe-network.
  • To use a cloud model provider instead, enter its key in Settings > Provider (for example an OpenAI key from your provider's console). Never commit .env or keys to git.

Limits and costs

  • Licence: LobeHub Community License, based on Apache 2.0 with extra conditions. Commercial use without changing the code is allowed; developing and distributing a modified (derivative) version needs a commercial licence from LobeHub. Read it before reselling a fork.
  • Small local models give simpler answers than large cloud models.
  • Not everything works offline: the community market and web search need internet.
  • Self-hosting is free; you handle updates and backups.

Clean up

bash
docker compose down        # add -v only if you want to delete the data too
docker rm -f ollama        # add: docker volume rm ollama   to delete the models

Chats and files live in ./data and ./s3_data inside the lobehub folder.

  • Docs (Docker Compose): https://lobehub.com/docs/self-hosting/platform/docker-compose
  • Ollama provider: https://lobehub.com/docs/usage/providers/ollama
  • Agents: https://lobehub.com/docs/usage/getting-started/agent
  • Topics: https://lobehub.com/docs/usage/agent/topic
  • Source code: https://github.com/lobehub/lobehub
  • Licence: LobeHub Community License, no SPDX identifier (https://github.com/lobehub/lobehub/blob/canary/LICENSE)
  • Releases: https://github.com/lobehub/lobehub/releases
  • Ollama image: https://hub.docker.com/r/ollama/ollama

--- Not affiliated with or endorsed by the LobeHub authors. "LobeHub" is used only to describe the software.

Liked this guide?

Subscribe on YouTube so you don't miss the next tutorial.

Subscribe

Weekly AI news + new tutorials

One email a week. Unsubscribe anytime.

Free. We send one email a week and never share your address. Privacy