66,977 stars · MIT from the repository sidebar. · v1.17.0, the release page reading '01 Oct 22:17' with no year and taken as 2026-10-01 because that date is in the past. ungh.cc gives pushedAt 2026-10-12T15:45:36Z. · Track this in Scout
One program that reads documents you add and answers questions about them, with the model running locally or hosted.
▶Repo detailsthe review · specs · pros & cons · install
What it does
You install the desktop application or run it in Docker. You create a workspace and add documents to it — PDF, plain text, Word files and more. It splits each document, turns the pieces into number lists called embeddings, and stores them so a question can find the relevant passages. You then chat with a model, and it answers using those passages. The model can be local, through llama.cpp-compatible files, Ollama, LM Studio or LocalAI, or hosted, through OpenAI, Anthropic, AWS Bedrock, Gemini, Mistral and others. The store for embeddings is LanceDB by default, with PGVector, Pinecone, Chroma, Weaviate and Qdrant as alternatives. It also runs agents that carry out multi-step tasks, connects to tools over MCP, runs scheduled tasks, has a programming interface, and supports speech. The newest release is v1.17.0, from 1 October 2026. What the desktop version does not have is multiple users with separate permissions, or the chat widget you embed in a website. Both exist only in the Docker version.
Why it matters
Who it suits. Anyone with a pile of documents they want to ask questions of and a reason not to upload them — contracts, research papers, internal notes, anything under a confidentiality agreement. It also suits a small team, because the Docker version gives each person their own account and controls what each can see. Skip it if one hosted assistant already handles your documents and the privacy is not a concern, because this is more to run for the same answers.
Verdict. The most complete single program in this space and the fastest route from nothing to asking questions of your own files. The desktop application is genuinely a double-click install, which almost nothing else here can claim. Two things to decide up front. First, pick Docker over the desktop version if more than one person will use it, because the desktop version has no accounts. Second, a local model needs real memory, and the project gives no figures, so expect to try one and find out. open-webui/open-webui at 154,337 stars is the larger alternative, covered in Edition 23, and is the better choice when chatting with models is the main job and documents are secondary. janhq/jan at 44,930 stars is the simpler desktop-only option when you want offline chat and nothing else.
- A desktop install for Mac, Windows and Linux, which is as close to one click as this gets.
- Works with local models or hosted ones, and six different stores for embeddings.
- Multiple users with separate permissions, when run in Docker, plus an embeddable chat widget.
- Multiple users and the chat widget are missing from the desktop version, which is the version most people install first.
- No memory, disk or processor figures are published, so a machine cannot be sized from the documentation.
- The long list of supported providers means a lot of settings to understand before the first good answer.
open-webui/open-webuiA larger interface built around chatting with models, at 154,337 stars, with code on 11 October 2026; Edition 23 covered it.
Track this in Scout- janhq/jan
A desktop application for running models entirely offline, at 44,930 stars, with code on 13 October 2026; simpler, with no document workspaces.
Track this in Scout
git clone https://github.com/Mintplex-Labs/anything-llm.git cd anything-llm/docker cp .env.example .env docker compose up -d
