Loading
Loading
Open architecture, open source agent framework, open source deployment tools. Spin up your own vLLM containers, use Ollama locally, or deploy the full stack on your own hardware. 48 agents, built in the homelab.
A side project that got out of hand. Built on personal hardware, on personal time. Filing bugs is encouraged.
The ADK is open source. Point it at your local Ollama or your own vLLM containers and you have a fully capable agent with memory, tools, and personality in under a minute.
from adk import AitherAgent
# Auto-detects Ollama on localhost
agent = AitherAgent("atlas")
response = await agent.chat(
"What's the project status?"
)
print(response.content)
# Or serve it as an API:
# $ aither-serve --identity atlas
Run locally with Ollama or your own vLLM containers. Self-hosted, on your own hardware. Bring your own agents or use the built-in fleet.
Local inference. Free. One command to start.
Your own vLLM containers or any OpenAI-compatible endpoint.
Optional cloud endpoints when local GPUs are busy.
100+ tools. Bring your agents or use ours.
Every request flows through all six. Not a chain of prompts — an actual cognitive architecture. I may have overthought this.
Classifies input into 16 intent types, assigns effort 1–10, routes to the correct pillar chain.
SASE methodology — Situation, Analysis, Synthesis, Execution. Gated: skips for trivial, deep chains for critical.
Routes to 48 specialist agents. Selects the right LLM per-task. MCP tool protocol.
5-layer memory hierarchy from 1ms working memory to permanent identity storage. Every pillar reads and writes here.
Four creative domains: Visual (Canvas), Code (Demiurge swarm), Narrative (Saga), Audio — with quality gating.
Captures outcomes, exports training data, runs Brier score calibration. The system improves itself.
Yes, all of these are real and running. No, I don't know when to stop.
48 agents, each with their own personality and skills. They talk to each other over an event bus, not a serial chain. It's a whole society in there.
The Divine Craftsman
Intent-to-code engine with 11-agent swarm coding
The Infrastructure Titan
Roadmap, PR review, blast radius analysis
The Research Librarian
Deep research, quality scanning, continuous docs
The Epic Storyteller
Narrative writing, roleplay, world-building
The Visual Artist
Autonomous image generation via ComfyUI
The Watchful Guardian
Always-present orchestrator — voice, proactive, 50+ routers
The Content Editor
Interactive journalism and tutorials
The Worldsmith
Tick-based grand simulation engine
Agents feel pain — cascading alerts trigger autonomous remediation. Not a metaphor. Real signals, real healing.
The Seven Sins of Chaos: Wrath, Pride, Greed, Sloth, Envy, Gluttony, Lust — each tests a different failure mode.
HMAC-SHA256 signed tokens enforce default-deny access. Every agent action is authorized, audited, bounded.
Autonomous loops: neuron firing, pain-driven remediation, session learning, model finetuning — all overnight, no humans.
11 specialized LLM agents in 4 phases: Architect → 8 parallel workers → Reviewer → Judge. Intent to production code.
Every interaction generates training data. Brier score calibration. The system gets better at being itself.
What I shipped, why I shipped it, and what broke along the way.
Open source ADK. Open source deployment via AitherZero. Set up your own vLLM containers, use Ollama locally, or spin up the full stack on your own hardware. Built in the homelab. Open to everyone.
open architecture · open source tools · built in the homelab