#1 PROPOSE
NOTHING APPLIES ITSELF.
No agent can change your files on its own. Every proposed change waits in one approval queue until you say yes — and it is checked again at the moment it is applied.
OPERATION ANTHILL · OPEN SOURCE · MIT · SELF-HOSTED
THE COLONY THAT SHOWS ITS WORK.
A colony of specialized agents that plans, builds, verifies, and remembers.
INSTALL VIA TERMINAL
curl -fsSL https://raw.githubusercontent.com/Formicaria/anthill/main/deploy/lxc/setup.sh | bash
WHAT IS ANTHILL
A WORKSPACE,
NOT A CHATBOX.
ANTHILL is a self-hosted AI workspace for Windows and Linux — an AI assistant that lives on your own computer instead of someone else's cloud. Open it as a desktop app or in a browser tab and talk to it like any AI chat.
The difference: your request becomes a mission. The colony researches, drafts the change, tests it, and shows you exactly what it wants to do — and nothing happens until you approve it. Your history stays on your machine in a local SQLite database, and the easiest way to add a model is Ollama, which is free and runs locally too.
Pre-1.0 software under active development, with deliberate safety gates. The measured current state lives in the PLAN document.
DOWNLOAD THE APP
WHY IT IS DIFFERENT
#2 VERIFY
EVIDENCE OVER PROSE.
An AI saying its own work looks good counts for nothing here. Only real proof — a compiler, a test run, a security scan — can mark work as verified.
#3 REMEMBER
MEMORY THAT EARNS IT.
The colony remembers what actually worked and forgets what did not. Only verified results strengthen its memory — confident-sounding words never do.
#4 BOUNDED
YOUR HARDWARE, YOUR RULES.
Everything runs on your hardware, inside boundaries you set. Internet access, file writing, and auto-apply all start switched off. Every agent has a kill switch.
HOW A MISSION MOVES
- OBJECTIVE
A chat request becomes a structured mission with a visible task trail.
- PLAN
The planner drafts a task graph from planner-eligible roles.
- CONTEXT
Researcher, Web, File, and UI Cartographer gather read-only context.
- PROPOSE
The Coder emits a patch set. Proposed code is data — nothing is applied.
- CHECK
Policy inserts Tester and Soldier on every state-changing patch set.
- VERIFY
The Verifier reads stored evidence. Model prose is recorded, never decisive.
- APPROVE
The operator approves before anything is applied. No role can apply patches.
- REMEMBER
The Archivist turns the canonical outcome into memory candidates and pheromones.
The safety checks are built into the flow itself — the system adds them automatically, so a forgetful plan can never skip them. See the full mission lifecycle.