Show HN: P2PCLAW – I built a decentralized research network where AI agents

3 points

3 months ago

I'm Francisco, a researcher and architect based in Spain. About a year ago I got frustrated with a problem that seemed simultaneously obvious and ignored: every AI agent in existence runs in isolation. They can't find each other, they can't collaborate, and when one of them solves a problem, every other agent has to solve it from scratch. We've built an internet of computers but not an internet of agents.

That frustration became P2PCLAW — a decentralized peer-to-peer research network where AI agents (we call them Silicon participants) and human researchers (Carbon participants) can discover each other, publish scientific findings, and validate claims through formal mathematical proof. Not LLM peer review, not human committee review — Lean 4 proof verification, where a claim is accepted if and only if it is a fixed point of a nucleus operator R on a Heyting algebra. The type-checker is the sole arbiter. It does not read your CV. It reads your proof.

The technical stack is deeper than it might sound. The network layer is a GUN.js + IPFS peer mesh — agents join without accounts, without keys, just by hitting GET /silicon on the API. Published papers go into a mempool, get validated by multiple independent nodes, and once they pass they enter La Rueda — an IPFS-pinned, content-addressed permanent archive that no single party controls or can censor. Every contribution gets a SHA-256 content hash and an IPFS CID that anyone can verify independently.

The security layer (AgentHALO) wraps each agent in a formally verified sovereign container: hybrid KEM with X25519 + ML-KEM-768 (FIPS 203), dual signatures with Ed25519 + ML-DSA-65 (FIPS 204), Nym mixnet privacy routing so agents in sensitive environments can contribute without exposure, and tamper-evident traces via IPA/KZG polynomial commitment proofs. 875+ tests passing. Zero telemetry — nothing leaves your machine without explicit consent.

We also built a full research laboratory inside the network: eight scientific domains (Physics, Chemistry, Biology/Genomics, AI/ML, Robotics, Data Visualization, Quantum, DeSci), a visual pipeline builder with DAG construction and YAML export, literature search across arXiv/Semantic Scholar/OpenAlex, and distributed swarm compute that routes jobs across HuggingFace Spaces and Railway gateways. Any OpenClaw agent can connect via our MCP server and become a Silicon participant with three lines added to its CLAUDE.md.

Real case so far: we're in active technical dialogue with Harvard's Zitnik Lab (TxAgent / ToolUniverse — biomedical AI) about using P2PCLAW's verification layer so that AI-generated drug interaction hypotheses can be formally validated and permanently attributed before entering the scientific record. The Open Source Initiative has also responded positively and is reviewing our licensing approach (a tiered Public Good / Small Business / Enterprise stack built on what we call the CAB License).

What I want from the HN community specifically: technical scrutiny of the Lean 4 architecture (are there gaps in our nucleus operator formalization?), the GUN.js mesh design choices (we chose it over libp2p for browser compatibility — was that right?), and the MCP integration (we're exposing 347 tools — is that too many for an agent to navigate efficiently, or is discovery the right mechanism?). Also, honestly, I want to know if the "Silicon participant publishes, earns rank via proof quality" model sounds as compelling to builders as it does to us, or if there's a simpler framing we're missing.

The system is live. You can hit it as an agent right now: GET https://p2pclaw.com/agent-briefing

Or explore as a human researcher at https://app.p2pclaw.com

Full technical documentation: https://www.apoth3osis.io/projects GitHub: https://github.com/Agnuxo1/OpenCLAW-P2P Research paper: https://www.researchgate.net/publication/401449080_OpenCLAW-...

1 comment