Subscribe
AI
5 min read

Self-Hosted Chat UIs for LLMs in Late 2026: Open WebUI vs LibreChat vs AnythingLLM vs LobeHub vs Jan

The inference engines get the headlines, but the layer most teams actually touch every day is the chat front end: the self-hosted web app that sits in front of Ollama, vLLM, or an OpenAI-compatible API and gives a company its own ChatGPT. Five projects dominate that layer in November 2026: Open WebUI, LibreChat, AnythingLLM, LobeHub (formerly LobeChat), and Jan. They are no longer converging on the same product. Open WebUI is turning into a workspace that hosts coding agents, LibreChat is buildi

Self-Hosted Chat UIs for LLMs in Late 2026: Open WebUI vs LibreChat vs AnythingLLM vs LobeHub vs Jan

The inference engines get the headlines, but the layer most teams actually touch every day is the chat front end: the self-hosted web app that sits in front of Ollama, vLLM, or an OpenAI-compatible API and gives a company its own ChatGPT. Five projects dominate that layer in November 2026: Open WebUI, LibreChat, AnythingLLM, LobeHub (formerly LobeChat), and Jan. They are no longer converging on the same product. Open WebUI is turning into a workspace that hosts coding agents, LibreChat is building enterprise agent plumbing, AnythingLLM is doubling down on documents, LobeHub is chasing consumer-grade polish under a restrictive license, and Jan is a desktop app patching the security holes that server-mode sidecars create.

This comparison is based on each project's GitHub repository, license file, and release notes, checked on November 19, 2026.

Where the category stands

The release notes tell the story better than any benchmark. Open WebUI's v0.13.4, published November 17, added "Tether," which turns a user's own computer into part of a deployment: local models from Ollama or LM Studio, MCP and OpenAPI tool servers, and terminal servers appear only for that user, with addresses and keys never leaving the machine. Earlier in the v0.13 series, the project wired coding agents such as Claude Code and Codex into the model picker, so an agent chat is just another "model" you select.

LibreChat moved the same direction from a different starting point. Its v0.8.9 notes (November 17) lead with Chat Projects (shared instructions and reference files), passkey sign-in, MCP Apps with deliberate tool approvals, and an Agent Channels beta that bridges agents to Slack with linked accounts and approvals.

Both point the same way: the chat box is becoming the place where agents run, tools get approved, and work gets scheduled. The differentiators are now licensing, deployment footprint, and admin controls.

The five projects, fact by fact

Open WebUI - 159,078 GitHub stars, the category leader by a wide margin. Latest release v0.13.4 (November 17, 2026). Deploys as a single Docker container or with a one-line uvx command; SQLite by default, PostgreSQL for scale. v0.12.0 added realtime voice calls that route questions to your own model, animated avatars, multi-factor sign-in, and per-group default models. The catch is the license. Open WebUI uses a custom license, not a standard open-source one: deployments may not alter or remove "Open WebUI" branding unless the installation has 50 or fewer end users in a rolling 30-day period, or the operator has written permission (that is where the paid enterprise license comes in). For an internal tool the constraint is cosmetic; for anyone white-labeling a client deployment it is a blocker.

LibreChat - 47,341 stars, MIT licensed, and the most enterprise-shaped of the group. The project now lives under the LibreChat-AI GitHub organization. Releases: v0.8.8 (October 1) and v0.8.9 (November 17, tagged prerelease). Strengths are breadth of provider support, agents with MCP tool integration, code interpreter, and an admin panel. The cost is the footprint: the default docker-compose starts six services (api, admin-panel, mongodb, meilisearch, vectordb on pgvector, and a separate RAG API). A team that wants MIT licensing and audit-friendly admin controls pays for it in moving parts; the solo user on a small VPS should look elsewhere.

AnythingLLM - 68,428 stars, MIT licensed, from Mintplex Labs. Latest release v1.20.0 (November 17). Ships as both a desktop app and a self-hosted Docker server. Its identity is document chat: workspaces of files with retrieval built in, and an OpenAI-compatible server mode. The v1.20.0 release is a good signal of direction: the browser extension grew from a clipper into a full chat sidebar with near-parity to the main UI, including the ability to right-click any page text or image and add it as context. If the job is "chat with these documents" for a small team, this is the shortest path.

LobeHub (formerly LobeChat) - 85,842 stars, the second-most-starred project here and the one that has invested most in interface polish and a large plugin/agent ecosystem. The repository has been renamed to lobehub/lobehub and the software is on a 2.x train (v2.3.4, November 18). Licensing deserves a careful read: since version 1.0 the project uses the "LobeHub Community License," Apache 2.0 with extra conditions. Running it commercially as an unmodified frontend or backend service is permitted; developing and distributing a derivative work commercially requires a paid license from LobeHub LLC. Self-hosting for internal use is fine; building a product on top of it is not, without a contract.

Jan - 45,839 stars, Apache 2.0 licensed, and the only project in this list that is primarily a desktop app (built on Electron, with llama.cpp-based local runtimes and an OpenAI-compatible local API server). Latest release v0.8.8 (October 28) is instructive: it is a security update that forces links in model replies to open in the system browser rather than inside the app window, and adds "Trusted Hosts" origin checks so arbitrary web pages can no longer call the local API server. That is the honest trade-off of desktop-first LLM tooling: your laptop becomes an inference endpoint, and someone's browser tab was able to knock on it until this release. Teams that need central admin, SSO, or shared knowledge bases will outgrow Jan; individuals who want local models with a decent UI are its audience.

How to choose

Comparison axis diagram

On licensing, the split is clean. LibreChat (MIT), AnythingLLM (MIT), and Jan (Apache 2.0) are permissive open source. Open WebUI and LobeHub both run custom licenses that are free to self-host but restrict rebranding (Open WebUI above 50 users) or commercial derivatives (LobeHub). If procurement or a white-label requirement is in play, start with the MIT pair.

On deployment weight, Open WebUI and AnythingLLM each run as a single container. LibreChat's compose stack has six services. Jan runs no server unless you enable its local API. LobeHub sits in between, designed around Docker Compose with optional cloud features.

On what each is actually for:

  • Shared company ChatGPT on your own GPUs, with the fastest release cadence: Open WebUI.
  • Multi-provider teams that need MIT licensing, per-user attribution, and MCP-based agents: LibreChat.
  • Document question-answering for a small team, desktop or server: AnythingLLM.
  • A polished, agent-and-plugin-heavy assistant you never resell: LobeHub.
  • A single person running local models on their laptop: Jan.

The pattern worth noticing

Two things stand out from the November 2026 releases. First, the interface layer is absorbing the agent layer. Open WebUI treats coding agents as selectable models; LibreChat ships agent-to-Slack channels and tool-approval flows. Nobody is content to be a chat box anymore, which means your "just a UI" decision is really an agent-runtime decision.

Second, the security posture of self-hosted front ends is still maturing. Jan's Trusted Hosts fix and Open WebUI's Tether design (keys and endpoints stay on the user's machine) are both responses to the same problem: these tools expose powerful model and tool access over HTTP, often with defaults that assume a trusted network. Whichever project you pick, treat the chat UI as a real internal service: put it behind your SSO, keep the API server bound to localhost, and read the release notes when security fixes ship.

Sources

AILLMself-hostingOpen Source