The self-evolving, sovereign AI operating system. It's the only AI home that writes its own tools, runs fully on your own hardware, and keeps growing — even after you're gone.
Flowork is a local-first AI operating system built around one idea: an AI that improves itself. It creates its own tools when it hits a gap, freezes what's stable so it never breaks, adopts new model/tunnel/voice providers as plug-ins you can pull out anytime, and carries a persistent brain that grows with every task. No cloud lock-in. No vendor you can't unplug. It runs on a $5 box, a laptop, or a bootable USB stick — and it's yours.
Heads up: Flowork ships with empty credential slots. Bring your own keys (or log in via OAuth) from the GUI — nothing is baked in.
GUI is the only source of truth.
Linux / macOS / WSL2
git clone https://github.com/flowork-dev/Flowork-OS && cd Flowork-OS
./start.sh # builds, then serves the GUI on http://127.0.0.1:1987Portable (no install — USB / any machine)
Grab flowork-portable-<version>.zip from Releases, unzip, and run start-flowork.sh (or Start-Flowork.command on macOS, Start-Flowork.bat on Windows). A sanitized starter-brain is already inside — it works out of the box.
Bootable appliance (VM / dedicated USB)
Flash flowork-os-<version>.iso (or .usb.img.gz) to a stick and boot. A whole AI OS, no host install.
| Command | What it does |
|---|---|
./start.sh |
Build + launch the full stack (router + agent + local model) |
open http://127.0.0.1:1987 |
The GUI — the single source of truth for every setting |
open http://127.0.0.1:2402 |
Router console — providers, models, usage, mesh, console log |
| chat mr-flow | Your flagship agent — talk to it in plain language, from the GUI or Telegram |
./stop.sh |
Stop everything cleanly |
- A real self-evolution loop. When mr-flow lacks a tool, it writes one, tests it, and promotes it — no human in the loop. Stable code gets frozen (hash + immutability) so future changes can never silently break the core.
- Plug-and-play, vendor-proof. Models, tunnels, TTS/STT, image, embedding, web-fetch — every third party is a registry plug-in. If a company disappears, you pull the file out and the core still builds.
Register / Get / List, nothing hardcoded. - Sovereign & local. Runs your own local model by default; any LLM provider (Claude, Antigravity/Gemini, OpenAI-compatible, Ollama) slots in. Your data, your brain, your machine.
- An agent colony. mr-flow orchestrates a team — scanners, coders, researchers, self-evolution judges — each isolated, each with its own memory.
- White-label, multi-OS, no hardcoded paths. Zero company branding in the code; rebrand freely. Linux, macOS, Windows, or a flash drive.
GUI (:1987) |
Telegram | CLI tools | |
|---|---|---|---|
| Chat mr-flow | ✅ | ✅ | ✅ |
| Configure everything | ✅ (source of truth) | — | — |
| Alerts / notifications | ✅ | ✅ | — |
| Run agents / triggers / schedules | ✅ | ✅ | ✅ |
| Topic | Where |
|---|---|
| Architecture & subsystems | docs/ |
| Providers & models | Router console → Providers / Models |
| Brain & memory | Router console → Brain |
| Agents & self-evolution | GUI → Agents / AI Studio / Self-Evolution |
| Security (Threat Radar, MITM, scanner) | GUI → Threat Radar · Router → MITM |
| Auto-update & release | docs/AUTO-UPDATE.md |
Flowork grows by plug-ins, not by breaking the core. Add a feature through a sibling extension / registry seam — never edit a frozen file. See the architecture docs before opening a PR.
Issues and PRs welcome. This is open infrastructure for a sovereign, self-improving AI — build on it, fork it, make it yours.
AGPL-3.0 — built to stay free and to keep every derivative free.