Best Self-Hosted AI Workspace: Comparison and Evaluation for Teams

By Kortix Blog · Published: September 20, 2026

Documentation reviewed: September 20, 2026. This is a documentation-based comparison of public repository READMEs, licenses, and official product documentation, not a hands-on benchmark, security certification, or legal opinion. Vendor and product copy is labeled as such. GitHub star counts reflect the date of review and will have changed.

TL;DR: The best self-hosted AI workspaces compared

The current SERP for "best self-hosted AI workspace" returns home-lab tool lists, vendor blogs, and community threads — none of which evaluate platforms on team-oriented criteria such as shared company memory, permission scoping, isolated execution, and human review gates. This guide fills that gap with a structured comparison of six self-hostable platforms judged on the dimensions that matter when a team, not an individual, chooses a workspace.

Kortix is the strongest documented fit when a team wants a self-hosted AI workspace that ties together shared company memory, multi-agent orchestration, role-based permissions, isolated per-session execution, and configurable human review gates — all through a git-based configuration model the team owns. Kortix's repository carries Elastic License 2.0 (source-available, not OSI open source). Open WebUI is the best fit for teams that want a feature-rich chat UI with RBAC and Ollama support under a permissive (branding-required) license. AnythingLLM is the strongest choice for document-first RAG workflows with MIT licensing. Dify suits teams building visual AI workflows and apps with low-code tooling. LibreChat excels as a multi-model chat interface with strong model switching. n8n is a workflow automation platform with AI capabilities, not a workspace per se, but commonly compared.

Comparison table: Self-hosted AI workspace platforms

Platform	License	Self-host model	Multi-model support	Team memory / knowledge	Permissions	Isolated execution	Human review gates	Best for
Open WebUI	Custom (branding-required, source-available)	Docker, pip, Kubernetes	Yes — Ollama, OpenAI-compatible APIs, any provider	Persistent memory per user; Channels for shared spaces; RAG with 9 vector DBs	Granular RBAC, user groups, LDAP/OAuth/SSO	Open Terminal (per-user containers, Enterprise); Computer (standalone agent)	Not documented as universal pre-action gate	Teams wanting a rich chat UI with RBAC and Ollama
AnythingLLM	MIT (OSI open source)	Docker, desktop app, cloud deployments	Yes — 30+ LLM providers including OpenAI, Anthropic, Ollama	Workspace-based document isolation; user memories (automatic + manual)	Multi-user via Docker: Admin, Manager, Default roles	Not documented (agents run inside the server process)	Not documented as universal pre-action gate	Teams needing document-first RAG with permissive licensing
LibreChat	MIT (OSI open source)	Docker, Docker Compose, cloud deployments	Yes — 20+ providers: OpenAI, Anthropic, Google, Azure, Ollama, OpenRouter	Conversation history search; file search; RAG API (separate repo)	Multi-user auth (OAuth2, LDAP, email); Admin panel with roles, groups, per-role permissions	Code Interpreter uses sandboxed execution (ClickHouse/code-interpreter)	Code approval: Ask/Allow/Deny per role for file writes and commands	Teams wanting multi-model chat with strong model switching
Dify	Dify Open Source License (Apache 2.0 modified)	Docker Compose, cloud, Enterprise (VPC/self-host)	Yes — hundreds of proprietary/open-source LLMs, model providers	Knowledge/RAG pipeline; workspace-level knowledge bases	RBAC; SSO/SAML (Enterprise); multi-workspace (Enterprise)	Agent sandbox (agents run commands, install software in isolated env)	Not documented as universal pre-action gate	Teams building visual AI workflows and low-code apps
n8n	Sustainable Use License (fair-code, source-available) + Enterprise License	Docker, npm, cloud	Yes — AI nodes: OpenAI, Anthropic, Google, plus any model via HTTP	Not a workspace; no native team memory; workflows connect to external data stores	RBAC (Enterprise); owner/member roles in community edition	Not documented per-workflow; workflows run in the n8n process	Human-in-the-loop approval nodes available in workflow designer	Teams needing workflow automation with AI capabilities
Kortix	Elastic License 2.0 (source-available, not OSI open source)	Docker Compose (self-hosted control plane)	Yes — any provider with your keys via in-stack gateway (Anthropic, OpenAI, Google, Groq, xAI, DeepSeek, etc.)	Shared company repo as memory (memory/ directory in git); versioned, diffable, grepable	Role-based: owner, admin, member, manager, editor; SAML SSO/SCIM (Enterprise key)	Per-session disposable Linux sandbox on its own branch; microVM (Platinum) or container (default)	Change requests for repo work; configurable action approval gates; merge deny-by-default for agents	Teams wanting an integrated AI Management System with owned config, shared memory, and review gates

What is a self-hosted AI workspace?

A self-hosted AI workspace is a software platform that you operate on your own infrastructure to give a team access to large language models (LLMs), shared knowledge, and AI agents within a controlled environment. Unlike hosted SaaS tools such as ChatGPT Work or Claude Cowork — where model calls, data, and configuration live on the provider's servers — a self-hosted workspace keeps every layer on hardware and storage the team controls.

A self-hosted AI workspace differs from an agent framework in scope. Agent frameworks such as LangChain or CrewAI provide libraries for composing agent chains but do not include a user interface, permission system, persistent team memory, or deployment infrastructure. A workspace includes all of those. It differs from a single-vendor chat tool (such as ChatGPT) in model choice and data ownership: a workspace lets you route model calls through several providers with your own API keys, and your team's conversations and documents remain on your infrastructure.

Six criteria separate a team-ready self-hosted AI workspace from a personal AI tool: shared company knowledge that persists across sessions and team members, multi-agent orchestration so agents can coordinate work, permission scoping so each user and agent sees only what it should, isolated execution so agent activity cannot contaminate the host or other sessions, human-in-the-loop review gates so no irreversible action happens unapproved, and model choice so the team is never locked into one provider.

Evaluation criteria for teams: What to look for in a workspace

The six criteria below are a proposed evaluation framework, not testing this guide performed.

Shared company knowledge (team memory). A workspace used by a team needs a store of facts, documents, and context that persists beyond a single conversation and is available to every authorized user and agent. Some platforms implement this as a RAG (Retrieval-Augmented Generation) pipeline that indexes uploaded documents. Others store memory as files in a shared repository. Without a shared knowledge layer, each user or agent starts from zero on every task.

Multi-agent orchestration. A team workspace supports multiple agents that can work in parallel, coordinate on multi-step tasks, and share context. Single-assistant chat interfaces lack this capability. The orchestration model — whether agents are manually assigned, automatically scheduled, or triggered by events — determines how much of the team's workflow can be automated.

Permission scoping. Role-based access control (RBAC) lets administrators define who can read, write, or administer specific workspaces, models, tools, and data. The depth of the permission model matters: per-resource permissions for agents and humans, group-level policies, and integration with corporate identity providers (LDAP, SAML, OAuth) separate a team workspace from a shared chat room.

Isolated execution. When an AI agent runs code, installs packages, or accesses the filesystem, that activity should happen in an environment the agent cannot escape. Isolation technologies range from containers (Docker) to microVMs (Firecracker-style sandboxes). Without isolation, a compromised or misconfigured agent can read or damage the host system.

Human-in-the-loop review gates. A workspace that can take real actions — write files, send emails, create database records — must offer configurable approval gates before those actions execute. The strictest model requires a human to review and approve every change. Weaker models allow automatic execution for low-risk actions and require approval only for sensitive ones. A workspace with no review gate is unsafe for production use by a team.

Model choice and data ownership. A self-hosted workspace should support multiple model providers through a bring-your-own-API-key model. The gateway that routes model calls should run on the team's own infrastructure so that prompts and responses never pass through a third-party proxy. Ownership of the data — conversations, embeddings, configuration files — should be unambiguous: the team holds the storage, not the vendor.

Platform-by-platform breakdown

Open WebUI: Self-hosted chat UI with strong RBAC, limited shared memory

Open WebUI is a self-hosted AI interface that connects to Ollama and OpenAI-compatible APIs. Its repository (152,600+ stars) documents installation via Docker, pip, and Kubernetes, and its feature set includes granular RBAC with user groups, LDAP/OAuth/SSO authentication, local RAG backed by nine vector databases, Channels for team collaboration, persistent user memory, web search, and plugin support via Filters, Actions, Pipes, Tools, and Skills. Repository

Open WebUI ships with a custom license that requires preserving "Open WebUI" branding in the interface unless the deployment qualifies for an exemption (≤50 users per 30 days) or purchases an enterprise license. The GitHub license detector flags it as "Other." License

For teams, Open WebUI's RBAC is well-documented — administrators define roles, groups, and permissions — but its shared memory model is per-user by default, not company-wide. Channels provide shared real-time spaces, and notes can be attached to chats, but there is no system-wide shared knowledge repository that accumulates across sessions organically. The Open Terminal and Terminals (Enterprise) features add per-user isolated containers for code execution, but these are enterprise-tier capabilities. Repository · Enterprise docs

Open WebUI suits small-to-medium teams that want a feature-rich chat interface with strong authentication and RBAC, operate on a single model backend (Ollama or OpenAI-compatible), and can accept the branding requirement or pay for enterprise licensing.

AnythingLLM: RAG-focused workspace under MIT, single-user by default on desktop

AnythingLLM is a self-hosted RAG application that lets teams chat with documents, run AI agents, and manage workspaces. Its repository (66,200+ stars) is MIT-licensed and documents support for over 30 LLM providers, built-in LanceDB (with optional PGVector, Pinecone, Qdrant), workspace-based document isolation, multi-user support (Docker only, with Admin/Manager/Default roles), MCP compatibility, a no-code agent builder, an embeddable chat widget, and automatic and user-managed memories. Repository · LICENSE

AnythingLLM's multi-user mode requires Docker — the desktop app is single-user. Team permissions exist through three roles (Admin, Manager, and Default), but the system is designed primarily around document-level access rather than resource-level permission scoping. Isolated execution for agents is not documented as a universal sandbox; agents run inside the server process. Repository

AnythingLLM is the strongest choice for teams whose primary need is document ingestion and RAG — ingesting PDFs, DOCX, audio, YouTube transcripts, and GitHub repos — and who want MIT licensing with no usage caps. It is a weaker choice for teams that need multi-agent orchestration, per-session execution isolation, or configurable human review gates.

LibreChat: Multi-model chat with strong model switching, weaker team memory

LibreChat is a self-hosted AI chat platform that unifies multiple AI providers in a single interface. Its repository (44,400+ stars) is MIT-licensed and documents support for Anthropic, OpenAI, Google, Azure, AWS Bedrock, Ollama, OpenRouter, DeepSeek, Mistral, Groq, and others. Features include AI agents with MCP support, a code interpreter (sandboxed via ClickHouse/code-interpreter), Artifacts (React/HTML/Mermaid rendering), conversation search, multi-user authentication with OAuth2 and LDAP, an admin panel for managing users/groups/roles, code approval controls (Ask/Allow/Deny per role), and OpenTelemetry observability. Repository · LICENSE

LibreChat's strength is model switching: users can change providers mid-conversation, and administrators can configure per-role model access through the admin panel. The code interpreter sandbox is isolated, and the code approval controls for file writes and command execution provide a review gate at the tool level.

LibreChat's weakness for team workspaces is shared knowledge. The platform offers conversation history search and a separate RAG API, but it has no system-wide shared memory or knowledge repository that persists beyond individual conversations and is automatically available to all agents and users. It also does not document per-session isolated execution for agents (the code interpreter is sandboxed, but general agent execution is not documented as isolated). Repository

LibreChat suits teams whose primary use case is multi-model chat with flexible model switching, code execution in a sandbox, and who can live without company-wide shared memory.

Dify: Visual AI app builder with RAG and workflow, evolving team features

Dify is an open-source LLM app development platform that combines a visual AI workflow builder, a RAG pipeline, an agent framework, and model management. Its repository (156,500+ stars) uses the Dify Open Source License (based on Apache 2.0 with additional conditions — it blocks unauthorized multi-tenant SaaS resale). Features include a visual workflow canvas, comprehensive model support (hundreds of proprietary and open-source LLMs), a RAG pipeline with document ingestion from PDFs, PPTs, and more, autonomous agents with a sandbox for running commands and installing software, RBAC (SSO/SAML in Enterprise), and observability through Opik, Langfuse, and Arize Phoenix. Repository · License

Dify's community edition (free, self-hosted via Docker Compose) is single-workspace. Multi-workspace, SSO, and advanced RBAC require the paid Enterprise tier. Agent execution uses a sandbox environment, but the isolation model (container vs. microVM) is not documented at the level of the community edition. The visual workflow editor is Dify's strongest differentiator: it lets non-developers build multi-step AI workflows with logic branches, tool calls, and knowledge retrieval without writing code. Repository

Dify suits teams that want to build and deploy AI-powered applications (chatbots, assistants, workflow automations) through a visual builder and can accept the Community Edition's single-workspace limitation or pay for Enterprise. It is a less natural fit for teams that need a self-hosted AI workspace as a day-to-day collaboration environment rather than an app development platform.

n8n: Workflow automation, not a workspace, but commonly compared

n8n is a workflow automation platform that gives technical teams the flexibility of code with the speed of no-code. Its repository (205,000+ stars) uses a fair-code license — the Sustainable Use License and n8n Enterprise License — meaning the source is visible and self-hostable but carries use restrictions. Features include 1,500+ integrations, native AI capabilities (AI agent nodes connected to OpenAI, Anthropic, Google, and open-source models), visual workflow building with JavaScript/Python custom code, RBAC (Enterprise), audit trails (Enterprise), and human-in-the-loop approval nodes within workflow design. Repository · License

n8n appears in "self-hosted AI workspace" searches because it can connect AI models to business tools through automation workflows — for example, routing a Slack message through an AI agent and writing the result to a CRM. However, n8n is not a workspace in the sense of the other platforms in this comparison: it has no shared team memory, no persistent chat interface, no user-facing AI workspace UI, and no built-in document management or RAG pipeline. It is a workflow engine that can power automation behind a workspace but does not serve as the workspace itself. Repository

n8n suits teams that need to glue AI models to business applications through automated workflows and already have a front-end workspace for user interaction. It is not a replacement for a workspace such as Open WebUI, AnythingLLM, or Kortix.

Kortix: Self-hosted AI Management System with shared company memory and review gates

Kortix is a self-hosted AI Management System that keeps agents, skills, company memory, and connectors in one git repository, runs each session in an isolated sandbox on its own branch, and lands work through a change request a human approves. Its repository (20,200+ stars) carries Elastic License 2.0 (ELv2) — source-available, not OSI-approved open source — which restricts providing the software as a hosted or managed service to third parties but allows internal use. Repository · LICENSE

Shared company memory. Kortix stores team memory as files in a git repository the team owns, under a memory/ directory. Every agent and session reads from and writes to the same shared repository, so knowledge accumulates across sessions and is automatically available to every team member and agent. The repository description states the memory is "plain files today, a system that compounds what it learns over time." Repository

Multi-agent orchestration. Kortix supports multiple agents running in parallel, each in its own sandbox, with shared skills and memory. Agents are defined as markdown files in the repository and can be assigned specific tools, connectors, and scopes. The repository's README says agents are "OpenCode agents with a scoped reach into tools — markdown at the baseline." Repository

Permission scoping. Kortix documents role-based permissions with built-in roles: owner, admin, member, manager, and editor. SAML SSO, SCIM directory sync, custom roles, groups, and audit read require an enterprise key on self-hosted instances. The self-hosting page states connector credentials are "brokered server-side and never enter the machine," so raw API keys are not exposed to agents. Self-hosting page

Isolated execution. Every session runs in a disposable Linux sandbox on its own git branch. The default provider runs containers; the Platinum provider runs microVMs. The repository says isolation is per-provider and each sandbox is "fully isolated" with the agent able to "install, run and break anything; only what it commits survives." Repository

Human review gates. Work lands in the repository only through a change request a human approves. Kortix documents configurable action approval gates and "merge deny-by-default for an agent" — meaning the agent cannot merge its own work without human approval. The self-hosting page says approval gates are configurable for the actions that matter. Repository · Self-hosting page

Model choice. Kortix routes every model call through a gateway running on the team's own infrastructure. The self-hosting page lists Anthropic, OpenAI, Google, Groq, xAI, DeepSeek, Mistral, Bedrock, OpenRouter, and ChatGPT/Copilot subscriptions as connectable providers — all with the team's own API keys. Self-hosting page

Kortix suits teams that want to own the entire AI workspace configuration in a git repository, need shared company memory across all sessions and agents, require per-session isolated execution, and want configurable human review gates. The Elastic License 2.0 restriction on hosted-service use means Kortix is not the right choice if the team plans to offer the workspace as a commercial service to external parties, but it is fully compatible with internal deployment for the team's own use.

[Note: Kortix's repository describes itself as "the leading open-source alternative to Claude Cowork and ChatGPT Work" and "open-source AI Management System." Those are vendor positioning statements, not independent verification. As noted above, ELv2 is source-available, not OSI open source, and this guide classifies it accordingly.]

How we tested and evaluated

This comparison evaluates platforms on published documentation only — repository READMEs, LICENSE files, official documentation sites, and self-hosting guides — reviewed on September 20, 2026. No platform was installed, benchmarked, or penetration-tested for this guide. Feature claims marked "documented" are attested in the project's own repository or documentation; they are not independently verified through hands-on testing.

Research methodology: For each platform, the authors read the repository README, the LICENSE file, the official documentation (docs.*.com), and the self-hosting/installation guide. Feature tables were populated with claims verifiable in those primary sources. Missing features are marked as "not documented" when the primary sources do not mention them — absence from documentation does not prove the feature is absent from the code, but it does mean the feature is not discoverable to an evaluator reading the project's own materials.

Limitations: GitHub star counts are volatile popularity signals, not quality or security metrics. License classifications are based on the text of each project's LICENSE file on the review date; license terms can change between releases. No performance, security, or scalability testing was performed. The "best for" column in the comparison table is an editorial assessment based on the documented feature set, not a vendor ranking.

Frequently asked questions

What is the best self-hosted AI workspace for teams?

The best choice depends on the team's primary need. For a comprehensive AI workspace with shared company memory, multi-agent orchestration, isolated execution, and human review gates, Kortix is the strongest documented fit (Elastic License 2.0, source-available). For a feature-rich chat interface with strong RBAC and Ollama support, Open WebUI has the most mature authentication and permission system. For document-first RAG workflows with permissive licensing, AnythingLLM (MIT) is the strongest option. Open WebUI Repository · AnythingLLM Repository · Kortix Repository

Is Open WebUI free and open source?

Open WebUI's source code is publicly visible and self-hostable, but its license is not OSI-approved open source. The project uses a custom license that requires preserving "Open WebUI" branding in the interface unless the deployment qualifies for a small-scale exemption (≤50 users per 30-day period) or purchases an enterprise license. Open WebUI License

What is the difference between a self-hosted AI workspace and an agent framework?

A self-hosted AI workspace includes a user interface, authentication, permission management, persistent storage, and deployment infrastructure — everything a team needs to operate AI day-to-day. An agent framework such as LangChain or CrewAI provides libraries for composing multi-step agent chains but does not include a UI, permissions, storage, or deployment tooling. Teams commonly use both: the framework inside the workspace. LangChain · CrewAI

Does Dify support self-hosting?

Yes. Dify's Community Edition self-hosts via Docker Compose with a single command. The Enterprise edition supports VPC and on-premises deployment with additional features including multi-workspace, SSO/SAML, RBAC, and audit logs. The Community Edition is free and single-workspace; the Enterprise edition is licensed. Dify Repository · Dify Self-hosting docs

Can I use my own API keys with these platforms?

The platforms in this comparison that support bring-your-own-keys models are Open WebUI (set API keys in environment or UI), AnythingLLM (per-provider keys in settings), LibreChat (per-endpoint keys via yaml config), Dify (model provider keys in settings), and Kortix (keys set as encrypted project secrets, injected at session boot). n8n supports API keys through credential nodes in the workflow designer. Kortix Self-hosting · AnythingLLM Repository

What is Kortix's license?

Kortix's repository is licensed under Elastic License 2.0 (ELv2), which is a source-available license, not an OSI-approved open-source license. ELv2 allows internal use, modification, and distribution but restricts providing the software as a hosted or managed service to third parties and prohibits circumventing license-key functionality. Kortix LICENSE · Elastic License 2.0

How does team memory work in AnythingLLM?

AnythingLLM provides two types of memory: automatic and user-managed. Automatic memory captures facts from conversations and stores them for the workspace. User-managed memory lets users explicitly record information. Memories are scoped per workspace, not company-wide, so knowledge in one workspace is not automatically available in another. AnythingLLM Repository

Does n8n include a shared team memory?

No. n8n is a workflow automation platform and does not include a shared team memory, persistent chat interface, or document management system. It can connect to external databases, vector stores, and AI models through its integration nodes, but it has no built-in knowledge layer. Teams that need both workflow automation and a workspace typically use n8n behind another platform such as Open WebUI or Kortix.

What is the difference between AnythingLLM and Open WebUI?

AnythingLLM (MIT license) is built primarily for document chat and RAG — ingesting PDFs, DOCX, audio, code repos, and web pages into a searchable knowledge base that an LLM queries. Open WebUI (custom license) is built as a general-purpose chat interface for LLMs with a focus on model management, user authentication, RBAC, and a broader plugin ecosystem. AnythingLLM is stronger for teams whose primary need is document understanding; Open WebUI suits teams that want a full-featured chat experience with user management. AnythingLLM Repository · Open WebUI Repository

Can I self-host LibreChat?

Yes. LibreChat self-hosts via Docker Compose. Its repository provides deployment configurations for Docker, Docker Compose, Railway, Zeabur, and Sealos. The README documents setup for HTTPS, reverse proxy, S3 storage with CloudFront, and LDAP/OAuth authentication. LibreChat Repository

Next steps: Choose your primary workload first

A self-hosted AI workspace is only as good as the workload it supports. Start by writing down the specific tasks your team will hand to the workspace — document analysis, automated reporting, code generation, CRM updates — then verify that the platform you choose licenses, deployment model, team memory, isolation, and review gates match those tasks. For a deeper dive into agent workflow reliability, read the reliable AI agent workflows guide; for the broader platform-vs.-framework decision, read the open source AI agent platform guide.