Proof Layer

Proof Layer

Cryptographic governance receipts for AI agents. Issued before the agent acts. Ed25519-signed. Hash-chained. Verifiable offline.

Built by WORLD999_LABS. Published on npm under @proof-layer/*. The legacy @veridocs/* package names continue to resolve and forward to the new ones.

What this repo contains

This is the public half of Proof Layer — the parts you install and run on your own machine:

The kernel (governance engine, Gauntlet adversary/judge pipeline, billing, dashboard) is closed-source and runs at prooflayer.world999labs.com. Open-source clients + closed-source server is the same pattern Stripe, Resend, and Vercel use — you get full transparency on what runs locally and what data leaves your machine, while the proprietary server logic stays protected.

Quick start

npx -y @proof-layer/mcp@latest

Drop into Claude Desktop's claude_desktop_config.json:

{
  "mcpServers": {
    "proof-layer": {
      "command": "npx",
      "args": ["-y", "@proof-layer/mcp@latest"],
      "env": {
        "PROOF_LAYER_API_KEY": "pl_live_...",
        "PROOF_LAYER_API_URL": "https://prooflayer.world999labs.com"
      }
    }
  }
}

Restart Claude. Eight pl_* tools appear in the picker. Get an API key at prooflayer.world999labs.com.

Migrating from @veridocs/mcp? No code changes required. The legacy package name continues to resolve, the legacy vd_* tool names continue to work alongside pl_*, and existing VERIDOCS_API_KEY / VERIDOCS_API_URL environment variables are still accepted. You can switch the package name and env-var names on your own schedule.

Why pre-execution governance?

Most "AI safety" tools run after the model has acted — toxicity filters, output classifiers, post-hoc audit logs. By the time the alarm fires, the agent has already sent the email, merged the PR, or executed the trade.

Proof Layer flips this:

1. Agent proposes an action (e.g. "send wire transfer for $50,000")
2. Proof Layer evaluates → returns signed verdict: EXECUTE / BLOCK / REVIEW / SHADOW
3. Receipt is written to a hash-chained, Ed25519-signed audit trail
4. Anyone with our public key can verify the entire chain offline — no trust in our infra required

A broken hash chain = tampering, immediately visible. A missing receipt = the agent acted without permission.

Verifying receipts offline

Receipts are designed to outlive us. Here's how to verify a bundle without ever calling our API:

import { verifyBundle } from "@proof-layer/verify";
import fs from "fs";

const bundle = fs.readFileSync("./receipts.ndjson", "utf8");
const publicKey = fs.readFileSync("./prooflayer-public-key.pem", "utf8");

const result = verifyBundle(bundle, publicKey);
console.log(result);
// { pass: 147, fail: 0, total: 147, chainBroken: false }

Public key is published at prooflayer.world999labs.com/v1/public-key and rotates on a published schedule.

Resources

• 🌐 Homepage: https://prooflayer.world999labs.com
• 🎬 Live demo (no signup): https://prooflayer.world999labs.com/demo
• 📖 MCP docs: packages/mcp/README.md
• 🔒 Privacy policy: PRIVACY.md
• 🐛 Report issues: GitHub Issues
• 💬 Support: [email protected]

License

MIT — see LICENSE.

The kernel and dashboard are proprietary and not included in this repo. The MCP server, verifier, and all client SDKs published under @proof-layer/* (and the legacy @veridocs/* names) are MIT-licensed.

Contributing

Pull requests, issues, and feedback welcome. See CONTRIBUTING.md.