Pfff. I've come full circle back to where I started.
Started with simple manual reasoning — what worked, what didn't, scores per command. Stupid simple. Worked great.
Then I got lost. Docker Swarm, PostgreSQL replication, HAProxy, NixOS driver fights — none of which I knew well before this project. I was so deep in making the infrastructure work that I forgot to give ARGOS the basic rules. Like, don't run on the carpet, you'll fall on your face. That kind of basic.
Result: a system that runs 24/7 but sometimes breaks what was already working because its "reasoning" has no foundation. It tries to philosophize about the entire context when all it needs to do is execute a command and check if it worked.
My fault entirely. All the small technical fires made me lose sight of the one thing that should have been the foundation from day one.
The plan: reasoning becomes a base skill — different per operation type. Linux admin, programming, network protocols, infrastructure analysis — each gets its own ruleset. Like military training: take these unpolished, chaotic recruits and drill basic rules into them until they operate with discipline. That was always the idea. I just forgot it while putting out fires.
Going quiet for a while after a few urgent fixes. Need to figure out how to build this reasoning visually — something I can see as a relational map, not text in a database. A demo video will be attached once I have something worth showing.
Hope this mess sparks something in your head too.
— DarkAngel
ARGOS is a self-hosted autonomous AI assistant built for real infrastructure management. Not a chatbot wrapper. Not a demo. A system that runs your homelab.
It handles Docker Swarm, NixOS configuration, PostgreSQL replication with automatic failover, disaster recovery, a skill tree with 100+ verified procedures, error pattern learning, and talks directly to Proxmox, UniFi, Home Assistant, and pretty much anything else in your homelab.
- 2-node Docker Swarm — permanent leader + worker
- PostgreSQL 17 with streaming replication and HAProxy auto-failover < 3 seconds
- Ollama (local GPU) for free operations — Claude for reasoning
- skills_tree — 100+ verified procedures loaded dynamically per context
- DEFCON monitor — automatic DB promotion on node failure
- error_patterns — learns from repeated errors, activates error codes after 10 occurrences
This repo contains safe public templates only:
setup.sh— bootstrap template for new nodesdocker/— Swarm stack + Dockerfile templatesskills/— generic whitelisted skills (NixOS, Debian, Docker)argos-setup.yaml.example— configuration template
The full ARGOS core (API, chat engine, UI, tools) is still in active development.
Actively developed. Already running 24/7 on real infrastructure. See ROADMAP.md for what's coming.
- Docker + Docker Swarm
- PostgreSQL 17 (via Docker)
- Claude API key (Anthropic)
- Grok API key (optional but recommended)
- Ollama (optional, for local model)
Made with love, spite, and pure determination.
- CPU: 4+ cores x86_64
- RAM: 8GB
- Storage: 40GB
- GPU: not required (Ollama skipped, API-only mode)
- OS: Fedora 42+, Debian 12+, Ubuntu 22.04+
Not yet tested at this spec. At minimum config, ARGOS relies entirely on Claude/Grok API. No local inference. Everything else should work — in theory.
- CPU: Intel Core i7 6th gen or equivalent
- RAM: 32GB
- Storage: 60GB SSD
- GPU: NVIDIA RTX 2060 (6GB VRAM)
- OS: Fedora 44, Debian 13
Ollama runs smaller local models at this spec. Good enough for daily use and testing. This is the laptop/workstation tier — solid for development and light production.
- CPU: 8+ cores (Ryzen/Intel 10th gen+)
- RAM: 32GB+
- Storage: 100GB+ SSD
- GPU: NVIDIA RTX 3080 10GB VRAM
- OS: NixOS 25.11 or Fedora 44 (tested)
- Future builds planned: Debian 12/13, Ubuntu 22.04/24.04, openSUSE Tumbleweed
- Community challenge: Linux From Scratch (LFS) — if you manage this, you deserve a medal and a pull request
Runs qwen3:14b locally without breaking a sweat. This is what ARGOS was built and tested on. Local inference handles all read-only ops, health checks, and routine tasks for free.
- Node 1 (Beasty): NixOS 25.11, RTX 3080, primary compute + DB + registry
- Node 2 (Hermes): Debian 13 VM on dedicated server, permanent Swarm Leader
- PostgreSQL streaming replication with HAProxy auto-failover < 3 seconds
- DEFCON monitoring on both nodes with automatic failover
Truly autonomous operation without Claude/Grok API requires a large local model capable of multi-step reasoning, tool use, and infrastructure decisions without hallucinating. Think RTX 6000 Ada (48GB VRAM) or a multi-GPU setup.
I am quite attached to both my kidneys, so this remains a future plan.
This is v1 alpha. It works. It runs real infrastructure 24/7. But it is not polished, not sanitized for every environment, and v2 will bring radical changes.
What changes in v2: complete knowledge architecture redesign (Bubble DB), security compartmentalization with local GPU-only processing for sensitive data, encryption between containers, advanced reasoning engine, PWA mobile app, and full database restructure.
If you install v1 now — migration tools will be provided for v2. No data loss.
This release was developed and tested on a private network (11.11.11.x range) with username darkangel.
Some paths and IPs are hardcoded in templates — replace with your own before deploying.
v1 = working, tested, real infrastructure. Not sanitized for generic use yet. v2 will have fully dynamic configuration with no hardcoded values.
If you see 11.11.11.111 or darkangel in any file — that's a placeholder for your own values.