CrewAI setup for FrankKi

Setup

uv add mcp

import os
from crewai import Agent
from crewai.mcp import MCPServerHTTP
from crewai.mcp.filters import create_static_tool_filter

frankki = MCPServerHTTP(
    url="https://mcp.frankki.app",
    headers={"Authorization": f"Bearer {os.environ['FRANKKI_API_KEY']}"},
    streamable=True,
    tool_filter=create_static_tool_filter(
        allowed_tool_names=["mcp_health", "address_validate", "letter_create_draft", "letter_preview", "shipping_quote", "order_status"]
    ),
)

agent = Agent(
    role="Mail preparation assistant",
    goal="Prepare reviewed physical mail without sending autonomously",
    backstory="A conservative operator that stops for human review.",
    mcps=[frankki],
)

1. 1.Use the structured configuration so the bearer token stays in a header and tools can be allowlisted.
2. 2.Keep send and approval-decision tools out of the agent until the crew has a non-model human approval boundary.
3. 3.Call mcp_health first and stop if the expected mode or capability is absent.
4. 4.Persist IDs, quote, maxCostEuros, and idempotency keys in deterministic application state rather than relying on agent memory.

Required FrankKi safety boundary

Validate the address, create a stable draft, inspect every preview page, obtain a fresh quote, set maxCostEuros, persist a stable clientOrderId, and obtain an explicit human decision before a live submission. Framework approval controls supplement FrankKi approval; model text never substitutes for the human decision.

Continue with canonical agent onboarding, the draft guide, send guide, and error recovery.