A sovereign LLM agent framework: multi-provider orchestration, plugin skills, 4-layer memory and tool use — runs locally in one command, deploys anywhere.
Not a wrapper around one API. RedAeye routes across providers with key pooling, circuit breakers and cost-aware policy — and its skills are knowledge the agent fully learns.
Groq, OpenRouter, OpenAI, DeepSeek, Together, DeepInfra, Fireworks, Mistral — sorted by cost, then power, then speed. Automatic failover + round-robin key pools.
Drop a plugin file and the agent learns the full skill. A meta-skill generates new skills from natural-language descriptions.
Working memory, token-aware pruning, rolling episodic summaries, semantic recall and user profile — cross-session persistence.
Metacognitive planning, skill activation and bounded tool-use rounds — with dict/object-agnostic tool-call normalization.
python main.py for the CLI. The same engine powers a Gradio chat UI and a FastAPI (/chat, /skills, /health).
Keys live in .env or host secrets — never in code. Offline mock mode lets everything run and be tested with zero keys.
One engine, three brains: routing, skills, memory.
Reproduced output from the offline smoke test — the full tool loop was exercised end-to-end.
Zero keys needed for the offline demo. Add a free key when you want live intelligence.
# 1. get the code (Files tab of this Space / GitHub) pip install -r requirements.txt # 2. run offline instantly (mock provider, full tool loop) python main.py --offline # 3. go live with a FREE key (Groq / OpenRouter) cp .env.example .env # paste your key, then: python main.py
Routing priority is explicit: free → most advanced → fastest.
| Provider | Type | Free tier | Get a key |
|---|---|---|---|
| Groq | Speed + free models | Yes | console.groq.com |
| OpenRouter | Aggregator (:free models) | Yes | openrouter.ai |
| Google AI Studio | Gemini direct | Yes | aistudio.google.com |
| Mistral | Direct + free tier | Yes | console.mistral.ai |
| Hugging Face | Inference providers | Yes | huggingface.co |
| OpenAI · DeepSeek · Together · DeepInfra · Fireworks | Direct / aggregators | Pay-as-you-go | see .env.example |
This box talks to a real model directly from your browser (OpenRouter). Your key stays in this tab — nothing is stored on any server.
Free keys: openrouter.ai/settings/keys (key starts with sk-or-) · Model: openrouter/auto · The key is cleaned of invisible characters automatically.
python main.py — terminal agent on your machine. Keys in .env. Optional: python installer.py for Typical / Advanced / Custom modes.
Everything in this repo is Docker-ready (Dockerfile + FastAPI/Gradio app on port 7860, endpoints /chat /health /skills). render.yaml included for one-click free hosting.
Run locally, then cloudflared tunnel --url http://localhost:7860 — free public HTTPS URL in seconds, perfect for a live walkthrough.