Skip to content

Runtime

Settings

Runtime readiness

secrets are reported as booleans only

Reading runtime…

Agent wallet

Base Sepolia · testnet

Testnet only. Paid calls settle test USDC on-chain after approval and cannot be reversed.

Spending limits

Hard ceiling per run. Enforced before an approval card is shown and re-checked when the stored approval is consumed.

Per-run ceiling

0.25 USDC

testnet

Override server-side with MAX_RUN_SPEND_USDC (0–5). Per-tool prices are fixed in the registry. There is no mainnet path.

Data & history

Run history and the activity ledger live in this browser's local storage. They are not account-synced and not durable server records. Pending approvals are held server-side in memory and expire after five minutes.

0 local run(s)

Secrets never enter localStorage, prompts, logs or the audit trail. API key values are never rendered.

Configuration

Create .env from .env.example. Names only:

  • GEMINI_API_KEY
  • GEMINI_MODEL
  • WALLET_PRIVATE_KEY
  • FACILITATOR_PRIVATE_KEY
  • MAX_RUN_SPEND_USDC
  • CRAWL_ALLOWED_HOSTS

Security notes

  • Approval ids are opaque, server-issued, single-use and expire after 5 minutes.
  • The client never receives the model/tool state needed to resume a paid run.
  • Denial cannot invoke a paid tool; approve re-validates the tool, price, intent and budget.
  • Research crawling re-validates DNS and host allowlists on every redirect hop.
  • Paid endpoints verify EIP-3009 signatures and reject replayed nonces.
  • Request bodies are bounded and the agent endpoint is rate limited per IP.

Demo-scale limitation: pending approvals are stored in process memory. A multi-instance deployment needs a shared store keyed by the same opaque id. This is stated in the README rather than hidden.