AnythingLLM
Own your intelligence — open-source, local-first AI workspace with RAG, agents, and multi-user workspaces.
About
AnythingLLM's pitch is in the tagline: stop renting your intelligence. It is a full AI workspace — desktop apps for Mac/Windows/Linux plus a Docker-deployable server — that wraps your own model runner (Ollama, LM Studio, llama.cpp, vLLM) or cloud APIs in document chat with citations, persistent memory, isolated workspaces, and an @agent mode that can browse the web, run SQL, save files, and use MCP servers. At 67,000 stars it is one of the most-installed open-source AI apps in existence.
Independent reviews consistently praise the same things: the workspace isolation model (each workspace gets its own docs, memory, model, and settings) and the background memory process that extracts facts from chats and re-injects them later. The caveats are operational: multi-user mode only works in Docker and can't be reverted, there is no native enterprise SSO, and a March 2026 CVE (CVE-2026-32617, unauthenticated endpoints on default installs) means you should update past 1.11.1 and set a password before exposing it anywhere.
What it can do
- Desktop + self-hosted server from one MIT-licensed codebase
- Chat over your documents with citations (built-in RAG collector)
- Persistent memory: background fact extraction re-injected into future chats
- Isolated workspaces — separate docs, memory, models, and settings per workspace
- `@agent` mode: web browsing, SQL, file ops, and MCP server tools
- Works with local runners (Ollama, LM Studio, llama.cpp, vLLM) and cloud APIs
Setup
- Download the desktop app (Mac/Windows/Linux) or deploy the server via Docker Compose.
- Pick your LLM: point it at Ollama/LM Studio for local inference or add an Anthropic/OpenAI key.
- Create a workspace and upload documents (PDF, DOCX, text, websites) — the collector ingests and embeds them.
- Chat with citations; invoke
@agentin any chat for tool use (web, SQL, files, MCP servers). - For teams: enable multi-user mode on the Docker deployment and set a password before exposing it.
Prerequisites
A machine with enough RAM/VRAM for your chosen model (or a cloud API key); Docker for the multi-user server deployment.
Use cases
Works with
Trust score — 9/10
Scored 9/10 (verified tier, metrics-v1). Evidence: 67,159 GitHub stars (GitHub API), MIT license, pushed Oct 15 2026; independent reviews corroborate the workspace model, the memory behavior, and the multi-user/SSO limitations; the CVE-2026-32617 record is public. Strengths: enormous adoption, genuinely open, the best local-first story in its class. Docked for: (1) CVE-2026-32617 (unauthenticated default installs ≤1.11.1) — patched, but set auth before exposing it; (2) multi-user is Docker-only and irreversible, no native SSO; (3) heavy RAM usage on large document collections. License: MIT.
★ 67,159 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 10.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 9: starScore from 67,159 stars on Mintplex-Labs/anything-llm
- · Maintenance 10: pushed 2026-10-15 (recency 10)
- · Openness 10: MIT + official site and docs verified
- · Manual 8: independent reviews corroborate the workspace model, memory behavior, and multi-user/SSO limitations; CVE-2026-32617 is public record (patched in >1.11.1); docked for the CVE history, Docker-only irreversible multi-user mode, no native SSO, heavy RAM usage
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.