open source 3-node multi-agent swarm
YT-AyoChat turns millions of YouTube Shorts views into loyal community. An autonomous, governed 3-node multi-agent swarm that authentically answers your viewers and cultivates community.
The attention economy pays you in views and keeps the relationship. Your best viewer is already three videos away.
Zero.
Clicks. Replies answered. Invitations delivered. The feed moves on before you can say subscribe, and everything you earned stays inside it.
dead traffic, at scale
yt-ayochat reads every reply against your own knowledge base. A Python RAG pipeline, not a canned autoresponder.
Guardrails first: screening, rate limits, a circuit breaker. Then the answer, in your voice.
$ ayochat run --pipeline governed-rag
You are indexing our lore. We are indexing your intent. In an attention economy built on noise, the only way to find your people is to let the model map your boundaries while you map its safety guardrails. We are studying each other.
inbound comment
“Where is that choreo transition from?”
agent reply + sovereign invite
“Studio rehearsal 4 at 0:15! Full breakdown dropping in our Discord community.”
One command in. Thousands of authentic replies out. Dead traffic just became your community.
where this belongs
A creator who answers comments to build a community is doing the platform's own job. That work should not require a third-party add-on, an API key of their own, and a browser tab nobody asked them to keep open.
Every comment-to-DM tool on the market today is scaffolding bolted to the outside of YouTube Studio. Creators paste tokens between dashboards, grant scopes they cannot audit, and hand the relationship they are trying to own to whichever vendor sits in the middle.
The retrieval, the governance and the reply already run on Google Cloud Platform. Gemini 3.7 Flash serves the model. Model Armor screens the prompt. Sensitive Data Protection redacts before generation. The pipeline never leaves the ecosystem.
Answer comments and send the invite automatically
proposed · Studio → Community → Automation
This should be a toggle in the creator dashboard, not a repository you have to find. Until it is, the repository is open.
The governance layer is the interesting part. Fork it, measure it with DeepEval, and push the policy further than this repo took it.
The listener, gateway, retrieval and guardrail stages are separate modules. Read the one you want to change before you change it.
71 unit and golden benchmark cases run against the versioned dataset. Get a passing baseline on your machine before modifying policies.
The SGP intercepts prompt injections and redacts personal data before generation. Regional model councils handle multilingual slang.
interactive demo · 3-node multi-agent swarm
Test the Lumi Architecture in real time. Paste an incoming YouTube comment in English, Spanish, Arabic, or Portuguese to observe the Supervisor, Perception Node (LLM Council Router), and Autonomous Hive synthesize a grounded, 1-sentence sovereign creator response.
# Enter any YouTube comment above and click "Run Swarm" to observe multi-agent perception, language routing, and sovereign hive response generation.$ git clone https://github.com/thanedouglass/yt-ayochat.git
$ cd yt-ayochat && python3 -m venv .venv && source .venv/bin/activate
$ pip install -r requirements.txt
$ python -m scripts.run_agent --query "that footwork transition at 0:15 was insane!"When non-English comments arrive (Arabic, Spanish, Portuguese), the Perception Node routes to an open-source regional model council hosted on Hugging Face & OpenRouter.
Evaluated continuously against the Golden Test Dataset using DeepEval and Pytest to prevent regressions and hallucination drift.
Continuous append-only synthetic memory logs live interactions for offline distillation without file concurrency locks.
| Type | Video ID | Inbound Comment | Intent & Energy | Lumi Sovereign Response | Alignment Score |
|---|
Pairwise Elo ratings computed across 250+ multilingual test matches comparing cultural resonance, slang fidelity, and boundary defense.
| Model | Gemini 3.7 | Llama-3-8B | CamelBERT | BETO | BERTimbau |
|---|---|---|---|---|---|
| Gemini 3.7 Flash | - | 78% | 82% | 85% | 84% |
| Meta Llama-3-8B | 22% | - | 58% | 61% | 59% |
| CamelBERT (Arabic) | 18% | 42% | - | 52% | 51% |
| BETO (Spanish) | 15% | 39% | 48% | - | 50% |
| BERTimbau (PT-BR) | 16% | 41% | 49% | 50% | - |