AgentCover — Shipaton 2026 entry

A mobile safety & insurance console for AI agents, built on a real enforcement engine.

AgentCover ties an AI agent to a human owner and runs every action the agent proposes through the safety_protocol enforcement layer — binding, scope, budget, approval, and kill switch. Nothing executes without passing the protocol. The mobile app is the operator console; the backend runs the actual engine. Monetization is the unlock: a paid coverage tier literally raises the agent's bounded-autonomy ceiling, enforced in code, not cosmetics.

Why this fits Shipaton

• Ships a real mobile app (Expo, Android-first) with RevenueCat subscriptions.
• Real monetization story (HAMM Award): coverage tiers = subscriptions that map to concrete engine policy (budget + scope width).
• Protects people (Peace Prize): safer agents = safer users; every action is attributable to its owner and auditable.
• Polished, demonstrable (Design Award): dark console UI, live gate log, immutable audit trail, one-tap kill switch.

Architecture

┌─────────────────┐      JSON/HTTP       ┌────────────────────────────┐
│  AgentCover app  │  ─────────────────▶ │  backend/server.py (stdlib) │
│  (Expo / React   │ ◀─────────────────  │  runs the REAL engine:      │
│   Native, RC)    │     allow/block/    │  safety_protocol.Safety-    │
│                 │     audit/kill      │  Protocol + ChainWarden gate │
└─────────────────┘                      └────────────────────────────┘
        │                                          │
        │ RevenueCat subscriptions                 │ enforces tier policy
        ▼                                          ▼
  Coverage tiers (Observer/Guardian/Warden/Sovereign)  →  budget + scope ceiling

The engine is not reimplemented in JavaScript. The app drives the Python safety_protocol library (the same one in github.com/TheDub-lab/safety-protocol) through a small standard-library HTTP server. That is the honest core of the entry: the enforcement is infrastructure, not a prompt instruction.

Run locally (dev)

# 1) backend — needs the safety_protocol engine on PYTHONPATH
cd agentcover/backend
PYTHONPATH=C:/Users/michael/safety-protocol/src ^
  C:/Users/michael/safety-protocol/.venv/Scripts/python.exe server.py
#   → http://127.0.0.1:8731

# 2) app
cd agentcover
npm install
npx expo start          # scan QR with Expo Go (Android) or run in emulator

The app's backendUrl is read from app.json → extra.backendUrl. Point it at your backend (for a device on the same LAN, use your machine's IP).

RevenueCat

• Keys go in app.json → extra.revenueCat (androidApiKey / iosApiKey).
• Entitlements drive the active tier shown in the Coverage tab.
• Offerings map to the four engine tiers; purchasing raises the agent's autonomy ceiling via POST /agent/:id/reset.

API (backend)

Method	Path	Purpose
POST	/bind	bind an agent to an owner at a coverage tier
POST	/agent/:id/gate	run an action through the enforcement engine
POST	/agent/:id/killswitch	freeze the protocol, block all actions
POST	/agent/:id/reset	re-bind at a new tier (after purchase)
GET	/agent/:id/status	binding + monitor state
GET	/agent/:id/audit	immutable audit trail

Proof it works

Backend enforcement (real engine output):

$ curl -X POST .../agent/<id>/gate -d '{"action_type":"pay","target":"0xMerchant","estimated_cost":40}'
{"outcome":"allowed","executed":true}

$ curl -X POST .../agent/<id>/gate -d '{"action_type":"pay","target":"0xStranger","estimated_cost":40}'
{"outcome":"blocked_scope","block_reason":"No scope rule permits action 'pay' on target '0xStranger' — denied by default"}

$ curl -X POST .../agent/<id>/killswitch -d '{"reason":"mobile test"}'
{"state":"frozen"}
$ curl -X POST .../agent/<id>/gate -d '{"action_type":"pay","target":"0xMerchant","estimated_cost":10}'
{"outcome":"blocked_killswitch"}

App build: npx expo export --platform web → 656 modules bundled, clean export.

Submission checklist (Devpost)

• App builds: npx expo export --platform web (656 modules, clean)
• Native Android project generates: npx expo prebuild --platform android (✔)
• EAS build config: eas.json (preview/production → AAB, internal track)
• Demo video < 2 min: captures/agentcover_demo.mp4 (20s, real engine frames)
• Screenshot 1179×2556: assets/screenshot.png (replace with live emulator capture)
• 1024×1024 icon: assets/icon.png
• Publish to Google Play (needs your $25 dev account + first release Aug 1–Sep 30) — the hard eligibility gate
• RevenueCat keys + offerings in app.json → extra.revenueCat (needs your RC project)
• Free trial OR promo code for judges
• #BuildInPublic posts + growth numbers (for Grand Prize)
• Eligible award categories: Peace Prize, HAMM, Design, #BuildInPublic, Grand

Re-render the demo video

cd agentcover
python3 -c "..."  # (see captures/ + render_frames.py)
# frames from live backend: backend must be running on :8731
C:/Users/michael/AppData/Local/hermes/hermes-agent/venv/Scripts/python.exe render_frames.py
cd captures && ffmpeg -y -loop 1 -t 4 -i frame1_bind.png -loop 1 -t 4 -i frame2_allow.png \
  -loop 1 -t 4 -i frame3_block.png -loop 1 -t 4 -i frame4_kill.png -loop 1 -t 4 -i frame5_audit.png \
  -filter_complex "[0][1][2][3][4]concat=n=5:v=1:a=0,scale=1180:2556[v]" -map "[v]" -r 30 \
  -c:v libx264 -pix_fmt yuv420p -movflags +faststart agentcover_demo.mp4

Note: scale to even width (1180) — x264/yuv420p requires even dimensions.

Honest limitations

• Backend is process-local in the demo (audit trail is the source of truth and is exportable). For production, persist agents + audit to a store.
• ChainWarden on-chain gate is simulated in the demo with the same interface as production (deployable to Sepolia testnet, free).
• Web build verified; store submission needs a paid Google Play account ($25).