Example: agent-preflight

Location: examples/typescript/src/agent-preflight.ts

Run:

export MISTRAL_API_KEY=sk-...
cd examples/typescript
npm run agent

Environment variables:

Variable	Required	Notes
MISTRAL_API_KEY	Yes	Mistral API key for mistral-large-latest
CKB_RPC_URL	No	Defaults to https://testnet.ckb.dev

What it demonstrates

A Mistral mistral-large-latest agent (via the Vercel AI SDK) is given a proposePayment tool. The tool:

1. Receives a recipient lock args and an amount from the agent’s reasoning
2. Fetches the live registry
3. Runs preflightCheck against the recipient
4. Throws if the recipient is blacklisted
5. Returns a payment proposal if the recipient is clean

The firewall check is inside the tool implementation — not in the agent’s system prompt, not in the agent’s reasoning, not optional. If the agent is manipulated into proposing a blacklisted recipient, the tool throws before any payment proposal is produced.

This models the correct way to gate an agent’s financial actions: enforce the constraint in the tool, where it cannot be bypassed by prompt injection or adversarial instructions.