Ancilo
Local AI made simple – a private chat on your own computer, a coding agent, a model API for any tool, and a worker that Claude Code and Codex hand tasks to.
🌐 ancilo.app
Ancilo runs open models (e.g. Qwen) on your Mac with llama.cpp. It picks a model that fits your computer, downloads it, starts it – and makes sure it never makes your computer unusable. Everything stays on your machine unless you explicitly turn on web search or set up a cloud model.
For everyone
- Set up in a few clicks – a step-by-step setup asks what you want to use Ancilo for and recommends one model that fits your computer's memory right now. One click loads it.
- A private chat – write, ask, translate, summarize; answers in Markdown; chats are kept.
- Web search, if you want it – off by default. Wikipedia (no account) or Google through Serper (your own key); by default Ancilo shows the search query and asks before anything goes out, and answers name their sources.
- Your computer stays usable – a slider from Eco to Maximum decides how much Ancilo may take. Models load only into memory that is free, idle models are unloaded, and Ancilo backs off when memory gets short or the computer hot. A status bar shows how the computer is doing and offers one-click help when it gets tight.
- Your documents – attach PDF, Word or Excel files to a chat, or make a folder a chat project: answers name the file and page, and such a chat stays on your computer.
- Tasks – give Ancilo something to do with your files: a table of your invoices, photos sorted by year, a report from your contracts. It works in a copy of the folder; you see every change and keep it, drop it or undo it.
- Build something – describe what you want; a coding agent works in a copy of the project, and you keep or undo its changes.
For developers
- Model API – OpenAI- (
/v1/chat/completions, /v1/responses) and Anthropic-compatible (/v1/messages) on 127.0.0.1, with a reliability pipeline that makes small models call tools correctly.
- Delegation – Claude Code and Codex hand self-contained sub-tasks (tests, renames, small fixes) to your local model through MCP: cheaper, private, not rate-limited.
- Coding agent – chat with the agent about a project, approve what it may do, review diffs, apply; an integrated terminal; "try again with another model".
- Several models – roles, routing per task kind, side-by-side comparisons, A/B tests in live use, a leaderboard from your own results.
- Search – a hybrid code index and a knowledge base the agents and the assistant use.
- API first – every function is an operation, available over REST, the CLI and MCP alike.
Status: early, released. Signed and notarized downloads for macOS on Apple Silicon; the app updates itself (with your click). The CLI and daemon also run on Linux – build them from source.
Install
Download Ancilo.dmg (macOS 13 or later, Apple Silicon), open it and drag Ancilo into Applications. Open Ancilo – the setup takes it from there. All versions: Releases · Changelog.
Install from source
Requirements: macOS on Apple Silicon, Rust (see rust-toolchain.toml), Node 22+, just, cargo-nextest.
git clone https://github.com/stefangrunert/ancilo.git
cd ancilo
just app-install # builds Ancilo.app, installs it to /Applications, links `ancilo` into ~/.local/bin
Open Ancilo from Applications – the setup takes it from there.
Quick start
The same from the terminal:
ancilo setup # a chat model that fits this machine + an embedding model
ancilo add hf.co/unsloth/Qwen3-4B-GGUF # or add a model of your choice
ancilo connect claude # let Claude Code delegate to Ancilo
ancilo ui # open the app in the browser
Documentation
Development
just build # release build of the `ancilo` binary
just verify # formatting, lints, all deterministic tests, app checks
just app-e2e # the app in Chromium and WebKit against a real daemon with fakes
just test-real # product checks with real small models (~10–15 min; downloads them once)
just bench-reliability # the reliability benchmark (~45 min) – when working on the pipeline
just release # build, sign, notarize and draft a release (maintainers)
just verify runs offline and without a GPU: every external dependency – Hugging Face, models, web search providers, Claude Code and Codex – has a deterministic fake. How the test tiers and releases fit together: Contributing.
License
Apache-2.0
