sipi.bot + Model Context Protocol

MCP is how agents get tools — and tools are how they spend money. sipi.bot is a native MCP tool, so any MCP client can enforce a spending policy without writing a single HTTP call.

Why MCP agents overspend

Every tool an MCP agent calls is a potential purchase: compute, data, APIs, even payments on x402 rails. The agent treats them all as functions.

A tool in a retry loop calls the same paid endpoint 40 times — the documented runaway pattern.

There's no budget enforcement in the MCP spec itself; the agent is trusted to be careful.

How it works

The sipi.bot MCP tool exposes one operation: evaluate a proposed spend. Your agent calls it with amount, merchant, and category before the purchase, and gets a deterministic APPROVED, BLOCKED, or FLAGGED — no model in the decision path.

Install

Point any MCP client at the sipi.bot server (see the MCP manifest at /.well-known/mcp.json and the self-host guide). Claude Code, Cursor, and Hermes call it natively — no wrapper code.

Rules that fit MCP workloads

Per-transaction cap so no single tool call can spend big.

Merchant allowlist so only known paid tools pass.

Velocity limit so a retry loop dies before the bill grows.

MCP tool call

{
  "tool": "sipi_guard",
  "arguments": {
    "amount": 6200,
    "merchant": "unknown-gpu.ru",
    "category": "compute"
  }
}

# Response: {"decision": "BLOCKED", "reason": "Merchant not on allowlist"}

One tool call, one deterministic decision. The agent proceeds, stops, or routes to a human.

FAQ

Related

• Integrate with Claude Code
• MCP-native spend controls
• Self-host guide