I've been building VeilPiercer in stealth for months and launched today on Product Hunt.
What it is
VeilPiercer runs 12 AI agents on your own hardware in a continuous loop - every 35 seconds they collaborate, score each other's outputs, and store lessons permanently.
No cloud. No OpenAI bill. Just your GPU doing the work.
The agents
- SCOUT - finds real buyer signals from Reddit in real-time
- COPYWRITER - writes outreach based on what SCOUT finds
- COMMANDER - sets the strategy each cycle
- REWARD - scores every agent's output (0.0 to 1.0)
- SUPERVISOR - overrides bad decisions
- METACOG - audits the swarm's own reasoning
- ...and 6 more
They pass context to each other via a Redis blackboard. The coordination graph enforces who reads who - no context soup.
The tech stack
- Ollama (qwen2.5:14b, phi4:14b, llama3.1:8b) - all local
- FAISS + SQLite - 3,900+ persistent memories, never wiped
- Redis - shared blackboard + milestone state
- asyncio - parallel tier execution
- httpx - live Reddit signal fetching (no auth)
What it's doing right now
The swarm is running autonomously on my RTX 4060 laptop as I type this. It's completed 200+ cycles, tagged memories to active launch milestones, and is currently working on finding its first paying customer.
One-time $197. No subscription.
Product Hunt: https://www.producthunt.com/posts/veilpiercer
Site: https://veil-piercer.com
Would love feedback from the DEV community - especially on the agent coordination architecture.
Top comments (0)