LobeHub

Last reviewed Aug 26, 2026 · 797 words

LobeChat is two products wearing one name. The docker one-liner everyone tries first stores every conversation in your browser's IndexedDB — clear the browser profile and your chats are gone, and nothing syncs between devices. The version in the screenshots, with accounts, cross-device sync, file uploads, and the knowledge base, is "server database mode", and it wants PostgreSQL, S3-compatible storage, and an identity provider. Decide which one you are deploying before you deploy anything, because moving between them later is a migration, not a toggle.

Client mode takes five minutes and 2 GB

For a single person on a single machine, client mode is genuinely easy — the catalogue rates it so, and the 81,994-star project has the deployment docs to match:

docker run -d --name lobe-chat -p 3210:3210 \
  -e OPENAI_API_KEY=sk-... \
  -e ACCESS_CODE=pick-something-long \
  lobehub/lobe-chat

Port 3210, done. ACCESS_CODE matters more than it looks: without it, anyone who finds the URL chats on your API bill. Budget the full 2 GB of RAM the catalogue lists — this is a Next.js application, not a static page, and it shows at container start.

The catch is where your data lives. Client mode keeps everything in the browser, which is fine right up until you want the same history on your phone, or you reinstall Firefox. Treat client mode as a stateless front-end for API keys, not a system of record.

Server mode is a real stack, not a bigger container

The lobehub/lobe-chat-database image is where LobeChat earns its "framework" billing and where the work lives. You need three external pieces: PostgreSQL (with the pgvector extension, because the knowledge base stores embeddings), an S3-compatible object store for file and image uploads, and an auth provider — the docs walk through options like Logto, Clerk, and generic OAuth. That is three services to run and back up before LobeChat itself starts.

My honest advice: if that list made you hesitate, you are a client-mode user, and that is fine. If you are provisioning this for a household or a small team, the server stack is a Saturday-morning project, and the payoff — every device, one history, shared knowledge bases — is the whole reason LobeChat exists.

It speaks to almost every provider, including your own

Provider breadth is LobeChat's strongest card. OpenAI, Anthropic, Google, Mistral, and a long tail of hosted APIs are first-class, and local models slot in by pointing it at an Ollama endpoint on your network. Vision input, voice conversation, and the plugin system (web search, code interpreters, and a marketplace of community assistants) work across providers where the underlying model supports them. In practice I keep one cheap hosted model for search-augmented questions and route everything private to a local model — one interface, two trust levels.

Note what self-hosting does not change: LobeChat is a client. Unless you run local models, your prompts still go to whichever provider you configured. Hosting the UI yourself controls the history and the account data, not the inference.

Where it sits against Open WebUI

The short version: Open WebUI grew up around Ollama and local-first inference; LobeChat grew up as a polished multi-provider chat app with a marketplace. If your models are 90% local, Open WebUI's tighter model management usually wins. If you juggle three providers and care about assistant presets, voice, and interface polish, LobeChat is ahead. I keep the long comparison in LobeChat vs Open WebUI, and the wider field is on the AI chat UIs category page — this space moves fast and has at least four credible options in 2026.

The costs nobody puts on the feature list

Two recurring gripes to price in. First, LobeChat releases frequently, and the database image occasionally ships migrations — read release notes before pulling latest on a server-mode install, and snapshot Postgres first. Second, the Apache-2.0 license is clean, but the assistant marketplace is community content: treat prompts you install like code you did not write, especially on an instance with plugin execution enabled.

What I'd do

Single user, one main computer: client mode behind ACCESS_CODE, ten minutes, move on with your life. Anything more: do server mode properly on day one — Postgres with pgvector, MinIO or any S3 endpoint you already run, one OAuth provider — because retrofitting sync onto browser-stored chats is the painful path. Wire in Ollama for anything you would not paste into a cloud form, keep one hosted key for the rest, and pin image versions with Postgres snapshots before upgrades. Run this way, LobeChat is the most polished chat front-end you can own outright.