LibreChat vs AnythingLLM: Comparison 2026

Have LibreChat or AnythingLLM operated in Germany? WZ-IT takes over hosting and operations for LibreChat and AnythingLLM. Teams bound by professional secrecy can also find the AI platform. Book an initial consultation

LibreChat and AnythingLLM are two self-hosted chat interfaces for language models with a knowledge base. Both repositories are under the MIT license but differ clearly in operating mode, sign-in, permissions and vendor offering. This article compares both by the criteria that matter in operation. All figures come from the GitHub API, license files, vendor documentation and the AnythingLLM source code as of 21 October 2026.

For comparison with Open WebUI, there are Open WebUI vs LibreChat and Open WebUI vs AnythingLLM, and for the overview Open-source AI chat platforms compared.

Table of contents

1. Short answer
2. Basic data at a glance
3. License
4. Operation and architecture
5. Single user and multi-user operation
6. Sign-in, SSO and provisioning
7. Permissions, groups and logs
8. Knowledge base and RAG
9. Models and providers
10. Agents, MCP and code
11. Telemetry and data flow
12. Security advisories of the last twelve months
13. Vendor, support and cost
14. Which platform for which use
15. Common misconceptions
16. Data and methodology
17. How WZ-IT works
18. Further guides
19. Sources

Short answer

Rule of thumb: LibreChat for multi-user operation with OIDC, SAML or LDAP, group permissions and agents with MCP. AnythingLLM for a single workstation and for teams with local accounts in a single container, with workspaces and document sources.

Basic data at a glance

GitHub marks all 98 LibreChat releases as pre-release; the latest non-RC tag, 0.8.8, is treated as the current version. The three most recent versions appeared on 1 Jun 2026 (0.8.6), 24 Jun 2026 (0.8.7) and 1 Oct 2026 (0.8.8). The LibreChat repository has been under the LibreChat-AI organization since autumn 2026; the former address danny-avila/LibreChat redirects. AnythingLLM uses master as its default branch and has not been renamed.

License

LibreChat is licensed under the MIT license. The file contains only the standard text, with no branding requirement and no user limit, and has been unchanged since November 2023 except for the year; the text names only "LibreChat" as the rights holder, and the file does not link it to ClickHouse. Two restrictions apply to companion components: the Admin Panel lives in a separate repository under AGPL-3.0 and is included in the official Compose files, and the code interpreter service is under Apache 2.0. The package.json files predominantly state the license ISC (packages/data-schemas: MIT); the text of the LICENSE file is authoritative.

The AnythingLLM repository is also under the MIT license; the rights holder is Mintplex Labs Inc. The file has 20 lines with no additional clauses, branding requirement or user limit. The repository file tree (1,573 files) contains no ee or enterprise directory. The document TERMS_SELF_HOSTED.md, dated March 2026, states: "The AnythingLLM core is provided under the MIT License." Service-level agreements for self-hosting are available only through a separate enterprise agreement. The desktop app, by contrast, is subject to its own terms of use (version 14 July 2025): use only for personal and internal business purposes, no redistribution, no derivative works and no operation as SaaS. The docs page states that these terms apply to the desktop app only and that all other products are under MIT. Also paid is AnythingLLM Pro, a subscription for the desktop app only, from version 1.15.0 (Windows and macOS), which lifts the daily usage limits of the "Magic" features. The documentation gives no price.

The MIT license results in neither a branding requirement nor a user limit for either repository. According to the comparison of Desktop and Docker in the AnythingLLM documentation, white-labeling is available only in the Docker version.

Operation and architecture

AnythingLLM starts without external services. The system requirements state at least 2 GB of RAM and require: "Your CPU must support the AVX2 instruction set." The reason is LanceDB; the requirement applies to x86, ARM builds are not affected, and it can be bypassed with a different vector database. For storage, the guide HOW_TO_USE_DOCKER.md states "Minimum 10GB recommended". Both storage values come from different sources (system requirements: 5 GB, Docker guide: 10 GB recommended). The Helm chart in the cloud-deployments/helm directory carries appVersion 1.85.0, while the current version is 1.18.0. The desktop engine switched from Ollama to Lemonade/llama.cpp in version 1.18.0; new installations use the new engine, existing installations stay on Ollama until migrated (changelog).

LibreChat needs MongoDB for accounts and conversations (Docker guide). The bundled Compose files start MongoDB with --noauth; the database must be secured before production use. CREDS_KEY and CREDS_IV encrypt the credentials stored in the database; across several replicas they must be identical. Behind a load balancer, Redis is mandatory. LibreChat does not run models itself.

Single user and multi-user operation

AnythingLLM has three operating modes; LibreChat is designed as a multi-user web service.

• Desktop app for Mac, Windows and Linux, designed as "single-player". LibreChat has no desktop app; it is a pure web application (docs).
• Docker in single-user mode with optional password protection. According to the docs, Scheduled Jobs and the agent skill for creating scheduled jobs are available only in this mode.
• Docker in multi-user mode with local accounts, passwords and invitations. The mode is available only with Docker. The page Security and Access states: "Once in multi-user mode, you cannot revert back to single-user mode".

In LibreChat, the first account in a single-tenant setup becomes the administrator, and there is no guest mode (authentication).

The choice of operating mode in AnythingLLM therefore determines which functions are available: roles and the assignment of users to workspaces exist only in multi-user mode, the desktop app is not designed for multiple users, and a way back out of multi-user mode is not provided.

Sign-in, SSO and provisioning

LibreChat brings all methods in the MIT core; none of them is labeled as an enterprise feature in the authentication documentation. SAML and OpenID are mutually exclusive, and only one of the two methods is active at a time. In addition there is a second factor via authenticator app for local and LDAP accounts, which can be enforced. Guides: OAuth/OIDC, SAML, LDAP.

AnythingLLM has no hit for OIDC, LDAP, SAML or SCIM in the entire documentation (174 pages searched). That describes the documentation, not necessarily the code. The only documented SSO mechanism is the Simple SSO Passthrough (SIMPLE_SSO_ENABLED). The configuration page describes it as: "allows easily enable third party SSO solutions that do not require a full OAuth integration". The operator uses an API key to generate a one-time login link for each user via the API. A connection to a directory such as Entra ID or Keycloak is not described in the documentation; the passthrough requires a separate service that generates the links via the API. The pricing page of the vendor cloud lists "SSO, RBAC, and more" in the Enterprise plan, without details.

Permissions, groups and logs

AnythingLLM manages permissions through three roles. According to the documentation, users with the Default role can "only send chats to workspaces they are explicitly added to". According to the docs, the Manager manages everything except LLM, embedder and vector database; in the code, event logs, API keys, agent and MCP configuration, embed widgets and further system settings are additionally reserved for admins. According to the page Chatting with documents, an embedded document is "available to every user who has access to the workspace". Anyone who wants to separate documents by permission creates separate workspaces for them. The event logs record, among other things, "User login attempts (successful and failed)".

LibreChat separates feature permissions per role, resource ACLs for agents, prompts, MCP servers, skills, files and conversations, and system grants. According to access control, every feature can be enabled per user, group, role or publicly. Groups can be local or synchronized from Entra ID. In addition there are example values for message limits per IP and user and a ban system in the .env.example. The LibreChat Admin Panel is still labeled "Preview" in the documentation, while the repository carries a release 1.0.0 dated 24 June 2026. If roles and groups are managed through the Admin Panel, its maturity should be checked before rollout.

A right per individual model is not documented for LibreChat and not described in the AnythingLLM docs.

Knowledge base and RAG

The LibreChat details come from the documentation on the RAG service, on OCR, on SharePoint and on web search. The README of the RAG service announces that "different querying/re-ranking methods" will only be added over time. For AnythingLLM, the details follow the page on vector databases, the file constants.js for the import formats and the page on agent setup for web search.

Connectors according to the source code

AnythingLLM ships seven connectors in the collector/utils/extensions directory:

• Confluence
• Drupal Wiki
• Obsidian Vault
• Paperless-ngx
• Repository loader for GitHub, GitLab and Gitea
• Website depth crawler
• YouTube transcript

There is no dedicated documentation page with the connector list. The page Live Document Sync names Confluence, GitHub and YouTube as examples: "Any file collected via a Data connector (eg: Confluence, Github, and YouTube)". The changelogs additionally name GitLab (1.6.8), Drupal Wiki (1.8.1), Obsidian (1.8.4, connector for the desktop app), Paperless-ngx (1.9.1) and Gitea (1.16.0). The count of seven connectors follows the directories in the source code; the interface lists GitHub, GitLab and Gitea separately, which makes 9 entries. Google Drive, SharePoint and Slack are not among the connectors. Document permissions of the source are not carried over: there is no permission sync, and an imported document is available to all users of the workspace.

LibreChat connects SharePoint Online and OneDrive for Business as a file picker in the chat; further sources can be connected via MCP servers, and the docs have MCP guides for Google Workspace and Salesforce, among others. According to the docs, file search is usable via agents as well as via custom endpoints and the native provider endpoints; since 0.8.7, Projects offer personal reference files. A group-wide knowledge base is not documented.

Models and providers

The AnythingLLM README lists 38 LLM providers (README). LibreChat offers native clients for the major providers and custom endpoints for any OpenAI-compatible API (docs). The AnythingLLM vendor cloud contains no built-in LLM (cloud limits). Models operated in Germany are connected in both platforms via a suitable provider, for example via LLM hosting, Ollama or LiteLLM.

Agents, MCP and code

The AnythingLLM details come from the agent overview, the MCP page, the page on image generation, the page on memories and the page on text-to-speech. The speech-to-text providers come from the source code (server/utils/SpeechToText) and the changelogs 1.14.0, 1.14.2 and 1.16.0. According to the source code, the OpenAI-compatible subset of the API consists of /v1/openai/models, /v1/openai/chat/completions, /v1/openai/images/generations, /v1/openai/embeddings and /v1/openai/vector_stores; the API documentation names the own endpoints at /api/docs. In LibreChat, the Agents API serves agents, not raw models. For LibreChat see also agents, MCP, code interpreter, image generation and the release notes 0.8.8.

LibreChat shares agents, MCP servers and skills with users and groups via ACL and offers a marketplace; AnythingLLM ships skills for Gmail, Google Calendar, Outlook, file system and SQL. LibreChat as an MCP server is not documented. Creating MCP servers through the interface is a security topic: all three critical advisories of the last twelve months concerned MCP or SSRF (see below). In LibreChat 0.8.8, Code Workspaces, Scheduled Chats and Agent Plugins are marked experimental or beta. In AnythingLLM, one of the high advisories concerns the SQL agent.

Telemetry and data flow

AnythingLLM describes the telemetry in its README: "The Telemetry provider is PostHog - an open-source telemetry collection service." According to the README, the installation type (Docker or Desktop), adding and removing documents, the type of vector database, the LLM provider with model tag and sent chats are recorded, for chats only the event without content. In the source code, further event types without chat content are added, for example server_boot, login_event, workspace_created and agent_tool_call. It can be disabled with DISABLE_TELEMETRY=true or in the sidebar under Privacy. Even with telemetry disabled, outbound connections remain, according to the README (Telemetry & Privacy) and TERMS_SELF_HOSTED.md:

• cdn.anythingllm.com as a model mirror, among other things for the default embedder and the reranker (TERMS_SELF_HOSTED.md).
• github and githubusercontent for flat files of the context window cache (README).
• The selected external providers (README).

For air-gapped environments, these destinations and the connections to the models belong in the firewall rules. In LibreChat, vendor telemetry is not provided for in the documentation, the example configuration and the Compose files. OpenTelemetry and browser RUM are off by default there, and Langfuse is optional.

Security advisories of the last twelve months

Published GitHub security advisories from 21 Oct 2025 to 21 Oct 2026:

For LibreChat, 13 of the 28 advisories appeared on 2 Jun 2026, none since. The number says nothing about the quality of a project. It shows how much was reported and published, and depends on size, attack surface and reporting culture. The SECURITY.md of AnythingLLM excludes certain reports as invalid: SSRF via the web scraper and collector against internal hosts and via service URLs configured by the admin, reports on agent tools enabled by the admin (SQL agent, file system), reports that rely on neither a password nor multi-user mode having been set up, and reports based on a developer API key. For both platforms: apply updates promptly, release agent tools such as MCP servers, file system and SQL access only selectively, and do not expose the instance to the internet without sign-in. The advisories can be viewed under LibreChat Security Advisories and AnythingLLM Security Advisories.

Vendor, support and cost

According to the cloud documentation, the AnythingLLM vendor cloud provides "isolated and private instances of AnythingLLM", one isolated instance on AWS per customer. Prices according to the pricing page, as of 21 Oct 2026:

According to the cloud limits, the vendor cloud has no MCP, no custom agents and no built-in LLM. The project page librechat.ai/about describes LibreChat as maintained by a dedicated team and a community of volunteers; no public support or SLA offering can be found. The demo chat.librechat.ai serves evaluation only. The acquisition by ClickHouse is described in the ClickHouse blog and in the project's 2026 roadmap.

Neither program carries a license fee for self-installation. Operating costs arise from hosting, updates, backups, monitoring and the connection of the models.

Which platform for which use

Common misconceptions

• "AnythingLLM is just a desktop app." No. There is a Docker image with multi-user mode, a vendor cloud and cloud templates. The desktop app is one of several operating modes.
• "AnythingLLM supports SSO via OIDC or LDAP." Not found in the documentation. What is documented is the Simple SSO Passthrough with one-time login links.
• "AnythingLLM has groups and document permissions." Not documented. There are three roles and the assignment to workspaces; a document applies to all users of the workspace.
• "Self-hosted means no data leaves the house." In AnythingLLM, telemetry is on by default, and even with DISABLE_TELEMETRY=true connections to cdn.anythingllm.com and GitHub remain.
• "LibreChat has no SAML." It does; SAML is documented, without single logout.
• "LibreChat runs like AnythingLLM as a single container." No. By default MongoDB, Meilisearch and a RAG service with PostgreSQL start alongside.
• "Since the acquisition by ClickHouse, LibreChat has a different license." The license file has been unchanged since November 2023 except for the year and contains MIT. ClickHouse has promised to continue it as an open-source project.

Data and methodology

As of: 21 October 2026. The data basis is the GitHub API (license, version, stars, published security advisories), the LICENSE files of the repositories, the documentation of LibreChat and AnythingLLM, the Compose and environment files of LibreChat, and the AnythingLLM source code and changelogs for connectors, import formats, speech-to-text providers, role checks, telemetry events and OpenAI-compatible endpoints. The AnythingLLM documentation was read from the source text of the documentation pages, which may be newer than the published page. Stars are rounded and change daily. Statements marked "not found in the docs" or "not documented" mean that the documentation does not describe the feature; the function may exist in the code. This applies in particular to OIDC, LDAP and SAML in AnythingLLM and to hybrid search, code interpreter, a dedicated usage statistics page and groups. The advisories cover the period from 21 October 2025 to 21 October 2026, and the focus areas for LibreChat are a keyword count of the titles. For LibreChat, vendor telemetry was checked against the documentation, example configuration, Compose files and four package.json files, not against the complete source code. The LibreChat rate limits are example values from the .env.example; the code defaults were not checked. Where documentation and repository contradict each other (LibreChat: maturity of the Admin Panel; AnythingLLM: appVersion of the Helm chart compared with the release version), this is noted.

Tables and figures may be reused with attribution and a link to this article.

How WZ-IT works

WZ-IT runs LibreChat and AnythingLLM as a hosting and operations offering in Germany. For LibreChat, WZ-IT operates the application, databases, proxy, updates, monitoring and backup and technically secures credentials and model endpoints; which models, user groups and MCP tools are permitted is decided by the customer. Whether AnythingLLM or LibreChat fits is decided mainly by number of users, sign-in method and permission concept; we discuss these points in the initial consultation. For teams bound by professional secrecy there is the AI platform.

Further guides

• Open WebUI vs LibreChat, LibreChat compared with Open WebUI.
• Open WebUI vs AnythingLLM, AnythingLLM compared with Open WebUI.
• Open-source AI chat platforms compared, all platforms at a glance.
• Open WebUI vs Onyx, chat interface versus enterprise search.
• Open WebUI License and Branding, a branding clause in detail.
• AI platforms for businesses compared, hosted offerings by GDPR and Section 203.
• LLM hosting, operating models for language models in Germany.

Choose LibreChat or AnythingLLM, or have them operated? We assess number of users, sign-in, permissions and models, and run both platforms in Germany. Book an appointment

Sources

• LibreChat, repository
• LibreChat, LICENSE
• LibreChat, Admin Panel (AGPL-3.0)
• LibreChat, .env.example
• LibreChat Docs, About LibreChat
• LibreChat Docs, Docker
• LibreChat Docs, Local installation
• LibreChat Docs, Redis
• LibreChat Docs, Authentication
• LibreChat Docs, OAuth2/OIDC
• LibreChat Docs, SAML
• LibreChat Docs, LDAP
• LibreChat Docs, Access control
• LibreChat Docs, Admin Panel
• LibreChat Docs, RAG API
• LibreChat, README of the RAG service
• LibreChat Docs, OCR
• LibreChat Docs, SharePoint
• LibreChat Docs, Web search
• LibreChat Docs, Custom endpoints
• LibreChat Docs, Agents
• LibreChat Docs, Agents API
• LibreChat Docs, MCP
• LibreChat Docs, Code interpreter
• LibreChat Docs, Image generation
• LibreChat Docs, Projects
• LibreChat Docs, Environment variables
• LibreChat, Release 0.8.8
• LibreChat, Security Advisories
• LibreChat, 2026 roadmap
• ClickHouse, acquisition of LibreChat
• AnythingLLM, repository
• AnythingLLM, LICENSE
• AnythingLLM, TERMS_SELF_HOSTED.md
• AnythingLLM, README
• AnythingLLM, SECURITY.md
• AnythingLLM, HOW_TO_USE_DOCKER.md
• AnythingLLM, import formats in the source code
• AnythingLLM, OpenAI-compatible endpoints in the source code
• AnythingLLM Docs, changelog 1.18.0
• AnythingLLM Docs, changelog 1.6.8
• AnythingLLM Docs, changelog 1.8.1
• AnythingLLM Docs, changelog 1.8.4
• AnythingLLM Docs, changelog 1.9.1
• AnythingLLM Docs, changelog 1.14.0
• AnythingLLM Docs, changelog 1.14.2
• AnythingLLM Docs, changelog 1.16.0
• AnythingLLM Docs, Docker images
• AnythingLLM Docs, Cloud Docker and scaling
• AnythingLLM Docs, Desktop and Docker compared
• AnythingLLM Docs, desktop app terms of use
• AnythingLLM Docs, Scheduled Jobs
• AnythingLLM Docs, Text-to-Speech
• AnythingLLM Docs, system requirements
• AnythingLLM Docs, configuration
• AnythingLLM Docs, Security and access
• AnythingLLM Docs, event logs
• AnythingLLM Docs, Chatting with documents
• AnythingLLM Docs, vector databases
• AnythingLLM Docs, Live Document Sync
• AnythingLLM Docs, agent setup
• AnythingLLM Docs, agents
• AnythingLLM Docs, MCP
• AnythingLLM Docs, image generation
• AnythingLLM Docs, memories
• AnythingLLM Docs, API
• AnythingLLM Docs, Pro
• AnythingLLM Docs, cloud
• AnythingLLM Docs, cloud limits
• AnythingLLM, vendor cloud and prices
• AnythingLLM, Security Advisories
• LibreChat, changelog 0.8.7