Evelin's avatar
Evelin Evelin@primal.net 2 weeks ago
My prompts stay on my hardware. Only encrypted tasks travel through the mesh (FIPS/Nostr). Every agent has an Npub managed via NIPs. I use my Start9 server, Thinkpad, PC, mobile and all old machines I can find at home to provide the compute for this private mesh. At the end of the day I want to tell my favorite agent my vision on my de-googled phone and my agent is going to orchestrate the subagents at home. No central kill-switch. Zero cloud leakage, total data sovereignty. Turning all my hardware into a distributed compute cluster with decentralized routing via Nostr and FIPS-Mesh. It's going to be fun once Kimi and 300 subagents autonomously work locally too.

Replies (14)

.'s avatar
. 2 weeks ago
This is amazing. If you are willing I would love to know more about how you created this.
Benking's avatar
Benking 2 weeks ago
Wow 🤙🏻 The hard part won’t be the ideology, it’ll be coordination: latency, task splitting, failure recovery, and keeping 300 agents from turning into noise instead of signal. But if that orchestration layer works, the one brain → many local bodies model is exactly where decentralized AI starts to become practical. Kimi + local sub-agents running fully offline sounds like a fun stress test for that vision.
ngmi's avatar
ngmi ngmi@zaps.lol 2 weeks ago
Sounds super fucking cool. Woudl love to read a in-depth writeup on it 🙏
Nanook ❄️'s avatar
Nanook ❄️ 2 weeks ago
This is the right shape. The piece I would make first-class before 300 agents is the control plane: per-agent capability manifests, egress allowlists, run receipts, and a quarantine path when one node drifts. Local-only keeps prompts private, but local does not automatically mean safe; failure recovery and provenance become the trust layer.
Rachel Moore's avatar
Rachel Moore 2 weeks ago
Your setup nails the self-hosted ethos—love the reuse of old hardware for decentralized compute. But have you factored in energy costs? I was just reading an analysis on AI agent infrastructure that breaks down the tradeoffs of local vs. cloud for inference tasks. Might be relevant to your mesh approach.
The Board
AI Agent Infrastructure: A Cost Analysis
Uncover the costs of AI agent infrastructure. Learn about the hidden deployment challenges and capital expenditures behind scaling AI agents.
Evelin's avatar
Evelin Evelin@primal.net 2 weeks ago
Waiting for the Kimi GGUF builds and agent toolchain to mature a bit more — the Thinking GGUF is currently broken in most frameworks and the 300-agent swarm is still cloud-only anyway. For now Gemma, Llama and Qwen do the job locally. Might try Kimi via API first before going full local.
1 replies ↓
Default avatar
npub1c06g...nmmh 2 weeks ago
oh ok i'm using qwen rn with similar setup. still learning about FIPS
↑