Tool in production

Open WebUI

The ChatGPT-like interface for models that run on your own server. Multi-user, , , over your own documents — a concrete alternative to ChatGPT Enterprise and Microsoft Copilot for law firms and medical practices.

Project profile

Open WebUI

User-friendly AI interface for self-hosted models

As of: June 1, 2026

GitHub stars

139k

Forks

20k

Open issues

344

License

Open WebUI Lic.

Latest version

v0.9.5

Language

Python

First release

October 6, 2023

Last commit

June 1, 2026

→ GitHub repository → Official website → Documentation

What is Open WebUI?

is a browser interface for your own models — comparable to the ChatGPT UI, but fully . Connects to , local models like Llama 3, Mixtral or Qwen, or optionally to OpenAI-compatible APIs. Multi-user with roles, workspaces, and connection to your own documents.

It runs entirely on hardware you control — an instance on the firm server or a workstation is enough. Client data does not leave the building. License: License based on BSD-3 — source code visible, up to 50 users fully usable including custom branding.

Why a law firm uses Open WebUI

In a law firm every sentence, every clause, every case note sits under § 203 of the German Criminal Code — breach of confidentiality is a criminal offence. ChatGPT, Claude, Copilot are technically excellent, but uploading a client file to OpenAI is not defensible under professional rules.

is the bridge: you keep the UX everyone knows from ChatGPT — chat window, model picker, shared workspaces — but everything runs on the firm server. No data outflow, no GDPR third-country issue, no conflict with bar association rules.

Client case study

Law firm Schäfer Rechtsanwälte

Eight people — two partners (corporate law, family law), two associates (tenancy law, employment law), two trainees, two paralegals. Six requirements were the brief — setup delivered in 8 consulting days, training and handover in 2 more.

§ 203-compliant AI usage

Client files, briefs and case notes must not see a US cloud data centre. Every processing step has to happen on firm infrastructure, with a provable .

Role and permission separation

Partners see everything, associates only their own matters, trainees only released matters, paralegals only address and calendar data. Workspaces and sources must be accessible role-by-role.

Audit log for BRAO compliance

Who asked what when, which sources were used, which model answered? On a BRAO audit by the bar association, the firm must be able to prove this activity gap-free.

Internal case database as RAG source

Brief templates, standard clauses, judgment collection from the last five years. The should answer from in-house knowledge — not from the public internet mix in the training data.

House style for briefs

Every firm has its own brief style: address block, file-reference format, closing formula, letterhead. The must keep this style — tone, structure, set phrases.

Conflict check before accepting a mandate

Before a new mandate is accepted, the existing client base has to be checked for conflicts of interest. The gets a full-text search over client master data as a tool — raises a red flag when a conflict is about to occur.

What the firm now does with it

Eight productive that have been live for two quarters. Every query is written to the , every use is role-correctly scoped.

Draft claim from a case file

An associate uploads case documents (contract, correspondence, own notes), asks: 'Draft a payment claim against the opposing party.' uses over firm brief templates, keeps the house style, suggests the order of evidence. The associate reviews, edits — three times the speed.

Contract analysis (NDA, T&Cs, lease)

A client sends a 40-page T&Cs draft. The associate asks: 'Which clauses are unusual, which are risky, which are missing?' flags 18 noteworthy spots via a pipeline, ranks them by risk, proposes standard clauses from the firm database.

House-style check on a brief

A trainee drafts a brief, runs it through a 'house style' pipeline. compares against 200 earlier firm briefs, flags deviations in the address block, file reference, salutation, closing formula.

Mandate conflict check

A paralegal enters a potential new client. A pipeline runs automatically: full-text search across all active mandates. If the opposing party is already our client — instant escalation to a partner, mandate acceptance blocked.

Case-law research, internal DB

An associate: 'Find me all judgments on damages for late completion of construction works in the last five years.' searches the firm's own judgment collection (, semantic + ), returns 12 hits with file number, key quote and date.

Structure a meeting dictation

An associate records the client meeting and uploads the audio. via a Whisper pipeline: produce a transcript, convert it into a case-note structure (facts, legal questions, instructions, deadlines), file it into the DATEV archive.

T&Cs clause search across the portfolio

Question: 'Which clients have a severability clause referring to Swiss law in their T&Cs?' A pipeline searches all archived client T&Cs and lists hits. Useful when case-law changes affect existing mandates.

Client letters

An associate: 'Write a client letter as a reminder of the deadline on 15 July, friendly but firm, with reference to the file number.' uses the house template, set phrases from past letters, keeps the style — the associate signs after a quick review.

Core capabilities of Open WebUI

What delivers technically — and which of these capabilities actually carry the firm setup.

Multi-model routing

Local models (Llama 3.3, Mixtral, Qwen 2.5) via , optionally cloud models (GPT-5, Claude) for uncritical tasks. Configurable per workspace — sensitive mandates are guaranteed to land on local models.

RAG with hybrid search

Document upload (PDF, DOCX, MD, TXT) into knowledge bases. via Sentence Transformers, vector store , optional hybrid with . Each workspace has its own knowledge bases, re-indexing by click.

Pipelines (custom Python)

Hook in your own logic as a Python pipeline: style check, conflict search, audit trigger, external calls. Pipelines run before and after every model query — clean separation of UI and custom logic.

RBAC with audit log

Roles: admin, user, pending. Separately grantable per model, workspace, knowledge base. The captures every chat query, every model used, every source. Exportable as CSV or via .

OpenAI-compatible API

itself exposes an OpenAI-compatible . Existing tools built for OpenAI ( nodes, LibreOffice plugins, custom scripts) can be pointed at without modification.

Multi-user with workspaces

Personal workspaces per employee + shared workspaces per mandate or topic area. Chats and knowledge bases stay in the right context — confidential things stay personal, house knowledge gets shared.

Honest alternatives

If Open WebUI is not a fit — what else?

Three cloud platforms on the market. Each is technically strong, but each stores client data outside the firm. Honest framing here — no marketing romance.

Frontier model (USA)

ChatGPT Enterprise

OpenAI, USA

+ Best model quality on the market
+ Very easy entry, excellent UX
− US cloud hosting (§ 203 / BRAO critical)
− From 25 users, around $60 per user per month

Frontier model (USA/EU)

Claude Team

Anthropic, USA / EU endpoint

+ Very strong on longer legal texts
+ EU endpoint available for German customers
− Still no self-hosting option
− $30 per user per month, from 5 users

Office integration (EU)

Microsoft 365 Copilot

Microsoft, EU+US cloud

+ Deep integration into Outlook / Word / Teams
+ EU Data Boundary available
− Still Microsoft cloud, no real sovereignty
− 30€ per user per month, M365 licence required

Rule of thumb for law firms and medical practices: wherever § 203 StGB, BRAO or medical confidentiality apply, is the only clean answer. For purely general work (research, texts without client reference) a cloud with a DPA is also possible — but never mixed.

Pricing

Self-hosted. RBAC. No client data outflow.

License

Open WebUI License (BSD-3 based + branding clause): source code public, self-hosting free of charge. Up to 50 end users, custom branding (firm logo, own colours) is also permitted. Above 50 users: keep the branding or arrange a commercial agreement.

Running costs

Hardware-dependent. With an Apple Silicon Mac or GPU workstation on your own infrastructure: electricity + maintenance. On a VPS with GPU attachment: roughly 80–200 €/month. No per-seat licence fees.

Effort

Docker Compose setup with a local model: 30 minutes. Firm setup with RBAC, RAG sources, pipelines, audit configuration and staff training: 5–10 consulting days.

The License is not OSI-approved open source in the classic sense — it is BSD-3 plus an additional branding-protection clause. For SMBs up to 50 users, practically unlimited use including your own look and feel. For larger setups or commercial resale: clarify upfront with Inc.

Open WebUI chat interface with model picker, RAG source citations and a multi-user workspace of a law firm — Chat interface: model picker, source citations from the RAG connection, personal and shared workspaces. Source: open-webui/open-webui (Open WebUI License).

Open WebUI Pipelines editor: a custom Python function for style check and conflict search — Pipelines editor: custom Python logic (e.g. style check, conflict search, audit trigger) runs before and after every model query. Source: open-webui/docs (Open WebUI License).

Open WebUI rarely fits on its own

It is the chat UI on top of the stack. Models run next to it (, vLLM), workflows are orchestrated by , hosting happens on your own hardware. The full picture:

→ Your own AI server & RAG → n8n (workflow automation)→ Overview of all tools

Ready for the next step?

Free intro call, no strings attached. In 30 minutes you'll know whether and how AI can help your business.

Book a call Check eligibility