The best self-hosted open-source AI workspace is one your team can own

Most self-hosted AI lists assume one person, one GPU and a free weekend. Your team needs shared memory, permissions, isolated execution and a way to approve work before it lands in production. Kortix, Open WebUI, AnythingLLM, LibreChat, Dify and n8n take different routes to those needs, and Kortix is the open-source AI Management System you can own.

What a team needs from a workspace

Shared team memory. A workspace earns its keep when what one agent learns is available to the next person and the next session. Ask where that memory lives and whether you can read, diff and back it up.

Permissions. You need to scope who can read, write and administer each project, and to hand each agent only the tools it needs. Identity integration through SSO, SCIM or LDAP matters once the team grows past a handful of people.

Isolated execution. When an agent installs packages and runs shell commands, that work belongs in a disposable environment it cannot escape. One shared container for every session is a security problem.

Human review gates. Any platform that can write files, send mail or touch a CRM needs a switch that stops it for approval. The strongest default is that nothing irreversible happens without a person.

Model choice. You should be able to point at Claude, OpenAI, Gemini or your own OpenAI-compatible endpoint with your own keys, per agent or per session, so a better model next quarter does not force a migration.

Kortix: one repo your company owns

Kortix keeps your agents, the skills they share, company memory, connector config and triggers in a single git repository you own, so the whole company is versioned and diffable (Kortix on GitHub).

Every session runs on its own isolated Linux machine, on a branch named after the session. The agent can install, run and break anything, and only what it commits survives. Work reaches the main branch through a change request a human reads as a diff, and merge is deny-by-default for an agent, so nothing lands unreviewed.

Kortix connects to 3,000+ apps plus any MCP, OpenAPI, GraphQL or raw HTTP API, and connector credentials are brokered server-side so the raw key never enters the sandbox (kortix.com). You pick the model per agent, per session or per message, with your own API keys, and you can self-host on a laptop, a VPS, your VPC or on-prem, or use the managed cloud.

Kortix is open source (Elastic License 2.0): self-host, read and modify the code.

Kortix is heavier to stand up than a single chat UI. If all you need is a private chat box over your own models, Open WebUI gets you there with less setup.

Open WebUI: the chat surface teams already run

Open WebUI is a self-hosted AI platform built to run entirely offline, and it connects to Ollama or any OpenAI-compatible API (Open WebUI on GitHub). It is strong at the chat layer: granular RBAC and user groups, LDAP, Active Directory, SSO and SCIM 2.0 provisioning, local RAG over nine vector databases, plus Channels and Notes beside the conversation.

Its memory is personal by default. Open WebUI remembers facts about you across your own chats, so what it learns does not become company knowledge. Open Terminal gives an agent a terminal and filesystem, but per-user isolated Terminals need the Enterprise tier, and configuration lives in the product database.

Open WebUI suits a team that wants a private, model-agnostic chat surface with serious identity and permissions, and does not need agent work to land as reviewed changes. It ships under the Open WebUI License, a permissive licence with a branding clause: the name stays visible unless a deployment covers 50 or fewer users in 30 days, holds written permission, or carries an enterprise licence (Open WebUI License).

AnythingLLM: document chat under an MIT licence

AnythingLLM is an all-in-one local-first app for chatting with your documents, with a no-code agent builder and MCP compatibility (AnythingLLM on GitHub). Its automatic and user-managed memories are scoped per workspace, multi-user support runs only on Docker, and the desktop app stays single-user. The project is MIT licensed (AnythingLLM LICENSE).

For a team whose main need is document ingestion and retrieval, AnythingLLM is a fast way to get there. It is a weaker fit when you need shared company memory across workspaces, per-session isolated execution, or a review gate before an agent touches a live system.

LibreChat: multi-model chat with approval controls

LibreChat unifies many providers in one chat interface, including Anthropic, AWS Bedrock, OpenAI, Azure, Google, Vertex AI, any OpenAI-compatible endpoint and Ollama (LibreChat on GitHub). It adds a sandboxed code interpreter, code approval controls that let admins set Ask, Allow or Deny for file writes and command execution, subagents with their own context windows, and an admin panel for users, groups and roles over OAuth2, LDAP or email login.

The gap for a company is shared knowledge. LibreChat's files and history stay per conversation, and there is no company-wide memory that every agent reads and writes. It is MIT licensed (LibreChat LICENSE), and it is a strong pick when flexible model switching and sandboxed code execution matter more than a persistent team brain.

Dify: build AI apps on a visual canvas

Dify is an open-source LLM app development platform with a visual workflow canvas, a RAG pipeline that ingests PDFs and slides, autonomous agents with their own sandbox, and support for hundreds of models (Dify on GitHub). It is licensed under the Dify Open Source License, an Apache 2.0 licence with additional conditions, so review those terms for your use case before you standardise on it.

The Community Edition self-hosts with Docker Compose, while SSO and RBAC sit in the Enterprise edition. Dify is the clearest option here for building AI applications and workflows.

n8n: automation behind the workspace

n8n is a fair-code workflow automation platform with native AI, a visual canvas, JavaScript and Python, and human approval steps (n8n on GitHub). It connects to 1,500+ integrations and runs self-hosted or in the cloud, under the Sustainable Use License and a separate n8n Enterprise License.

n8n is a workflow engine. It has no shared team memory and no chat UI, so it works best as the automation layer that moves data between the tools a workspace already uses. Pick n8n when you need to glue systems together behind the workspace your team actually uses.

Side by side

The six platforms differ most on where knowledge lives and how work gets approved.

When to pick which

Pick Kortix when you want the company's agents, skills, memory and connector rules in one repo you own, every session on its own isolated Linux machine, and each change landing as a diff a human approves. You choose the model per agent with your own keys, and self-host on a laptop, a VPS, your VPC or on-prem, or use the managed cloud. The free tier costs nothing to start, and Team runs $40 per seat per month, checked October 2026 (Kortix pricing).

Pick Open WebUI when the work lives in conversations and you want a private, model-agnostic chat surface with enterprise identity, and you do not need agent work to land as reviewed changes. AnythingLLM fits document-heavy teams, LibreChat fits multi-model chat with sandboxed code, Dify fits teams building AI apps, and n8n is the automation layer behind whichever workspace you choose.

If you are still weighing categories, read what an open-source AI agent platform means for a team and how to build reliable AI agent workflows. For the head-to-head view, see Kortix vs Open WebUI.

Free to start, free to self-host. Get started with Kortix.