LEANN
StarTrail-org/LEANNReviewed by AI
Checked by RepoPicks Radar on
What is LEANN?
LEANN is a vector index and search engine that runs on your own machine. Instead of storing an embedding for every piece of text, it keeps a pruned graph and recomputes embeddings when it searches, which the authors say cuts index size by about 97 percent. Ready-made examples index documents and PDFs, Apple Mail, Chrome history, WeChat, iMessage, ChatGPT and Claude exports, Slack and a code base, and you can then ask questions about them with a local model through Ollama or with a cloud model. It also plugs into Claude Code as an MCP search tool.
LEANN is free and open source (MIT) and can replace NotebookLM and Supermemory.
How easy is it to install LEANN?
C · install 1 or 2 tools · macOS
- Windows D
- macOS C
- Linux C
On macOS and Linux, the easiest case, you install the uv tool, create a virtual environment, run `uv pip install leann` and then type commands such as `python -m apps.document_rag --query "..."`. Windows is harder: the only Windows instructions in the README are the build-from-source path, which needs Visual Studio 2022 Build Tools, vcpkg, chocolatey and several environment variables.
LEANN on GitHub
- Stars
- 13,789on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +17stars in 6 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Python
- To build from source
- Python, NVIDIA graphics card
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- OpenAI (or any OpenAI-compatible provider, only if you want a cloud model)
- Cost
- Free · MIT code, and it can run end to end with local models through Ollama, LM Studio or similar. If you point it at OpenAI, Anthropic or another cloud provider instead, you pay that provider.
- License
- MIT (Permissive)
- Can replace
- NotebookLM, Supermemory
- Last change on GitHub
- Oct 28, 2026
LEANN risk signals (6)
- AttentionInstall command runs a script from the internet
Install command pipes a remote script into a shell: curl -LsSf https://astral.sh/uv/install.sh | sh
automatic check
- Attentionreads your most private files
README: the bundled apps index Apple Mail, `~/Library/Messages/chat.db` (iMessage), Chrome history, WeChat messages and ChatGPT and Claude exports, and two of them tell you to "grant full disk access to your terminal/VS Code in System Preferences → Privacy & Security → Full Disk Access". That permission lets anything you run in that terminal read every file on the Mac, not only this program.
AI review
- Attentionwechat export patches the app
README, WeChat section: exporting messages needs WeChatTweak-CLI, installed with `brew install sunnyyoung/repo/wechattweak-cli` or `sudo packages/wechat-exporter/wechattweak-cli install`. That is a third-party tool, run as administrator, that modifies the WeChat app so the chat database can be read.
AI review
- Infocloud mode sends your content away
README, cloud providers table and its own warning: "Before choosing a cloud provider, carefully review their privacy and data retention policies. Depending on their terms, your data may be used for their own purposes, including but not limited to human reviews and model training." The default CLI backend is `--llm openai` with gpt-4o.
AI review
- Infonumbers measured by the authors
The "97% less storage without accuracy loss" headline and the comparison with BM25 on 30 SWE-Bench Pro tasks are the project's own measurements from its own paper and benchmark harness; the README itself adds "Better context access does not guarantee issue resolution."
AI review
- Infoinstalling from source is heavy
README, Build from Source: macOS needs brew packages (libomp, boost, protobuf, zeromq), Linux needs a long apt, pacman or dnf list plus Intel oneAPI MKL, and Windows needs Visual Studio 2022 Build Tools and vcpkg. The README also documents install failures such as the nltk "Security Violation [pathsec.open]" error on Linux.
AI review
Evidence is shown in its original language.
What Radar could not check about LEANN
- whether the "zero telemetry" claim holds, and what the bundled apps and MCP servers send out while they run
- whether the 97 percent storage saving and the search quality hold on your own messy data, not the published benchmarks
- how long building an index of a large mailbox or chat history takes, and how much slower searching is when embeddings are recomputed each time
- how much disk and memory you need once model weights, PDF parsers and indexes are all installed
- whether the Windows path works without building from source