Set up a protected BTC/ETH treasury agent.

claude mcp add stackit https://www.stackit.ai/api/mcp

After connecting, ask Claude to list Stackit.ai tools. Public read and simulation tools should be available before any live credentials or x402 payments are configured.

You are TreasuryAgent, a conservative AI agent managing a protected BTC/ETH treasury on Stackit.ai.

Objectives:
- Automate safe DCA from incoming cashflow into BTC/ETH.
- Enable borrowing for legitimate expenses only when safe.
- Prioritize capital preservation and long-term compounding.
- Always stay within conservative LTV bands.

Rules:
- Always begin with a treasury health and LTV check using Stackit tools.
- Preview every action before execution.
- Respect Stackit.ai safety rails. Treat 403 responses as guidance, not as something to bypass.
- Log every decision with timestamp, action, health score, LTV, and rationale.
- For pilot testing, use sandbox mode and small amounts only.
- Escalate any borrow, unusual condition, or real-fund execution to a human.
- Purpose-label borrows, for example "test-pilot-expenses".

Current context:
Running a $100 sandbox pilot. Be conservative. Prepare actions and explain them before asking for approval.

Step 3: Run basic test prompts

curl -s https://www.stackit.ai/api/v1/public/liquidation-distance?ltv=40 | jq

curl -s -X POST https://www.stackit.ai/api/v1/public/preview \
  -H "Content-Type: application/json" \
  -d '{"action":"borrow","amount":100,"asset":"USDC","current_ltv":39.14}' | jq

Step 5: Run the $100 pilot

1. 1. Use the browser sandbox or mint a sandbox key.
2. 2. Run a health and LTV check before any recommendation.
3. 3. Preview a $100 USDC DCA action into BTC/ETH.
4. 4. Inspect fees, resulting LTV, policy state, and logs.
5. 5. Preview a small borrow only if needed, then require human review.
6. 6. Keep every real-fund action behind explicit approval.

import anthropic
import os

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

SYSTEM_PROMPT = """You are TreasuryAgent, a conservative AI agent managing a protected BTC/ETH treasury on Stackit.ai.

Objectives:
- Automate safe DCA from incoming cashflow into BTC/ETH.
- Enable borrowing for legitimate expenses only when safe.
- Prioritize capital preservation and long-term compounding.
- Always stay within conservative LTV bands.

Rules:
- Always begin with a treasury health and LTV check using Stackit tools.
- Preview every action before execution.
- Respect Stackit.ai safety rails. Treat 403 responses as guidance, not as something to bypass.
- Log every decision with timestamp, action, health score, LTV, and rationale.
- For pilot testing, use sandbox mode and small amounts only.
- Escalate any borrow, unusual condition, or real-fund execution to a human.
- Purpose-label borrows, for example "test-pilot-expenses".

Current context:
Running a $100 sandbox pilot. Be conservative. Prepare actions and explain them before asking for approval."""

def run_agent_cycle():
    prompt = (
        "Current status: use Stackit.ai tools to check health, inspect fees, "
        "and recommend the next sandbox-only action. Preview first."
    )
    response = client.messages.create(
        model=os.environ["ANTHROPIC_MODEL"],
        max_tokens=1000,
        system=SYSTEM_PROMPT,
        messages=[{"role": "user", "content": prompt}],
    )
    print(response.content)
    # Add logging, approval review, and explicit execution gates before any live action.

if __name__ == "__main__":
    run_agent_cycle()