Send iMessages from OpenAI Agents

What does the OpenAI integration actually look like?#

Function calling is the whole trick: give the model one tool that posts to the Blue Reacher API, and "text the prospect" becomes an action on your dedicated line.

from openai import OpenAI
import requests

client = OpenAI()

def send_imessage(phone: str, message: str) -> dict:
    """Send an iMessage. Human-approved outreach only."""
    r = requests.post(
        "https://api.bluereacher.com/v1/messages",
        headers={"Authorization": "Bearer brk_live_your_key"},
        json={"to": phone, "message": message},
    )
    return r.json()

• Register send_imessage in your tools array and the model calls it with a phone and a message. The response carries a message id, so the agent can confirm delivery with one more call.

Does ChatGPT Agent Builder connect without code?#

ChatGPT Agent Builder speaks MCP, and Blue Reacher hosts an MCP server at https://api.bluereacher.com/v1/mcp. Add it as an HTTP server with your bearer key and the messaging tools appear in the builder.

• The same hosted server works in Claude, Cursor and any MCP-capable client, so one key covers every agent surface your team runs.

• lookup_imessage checks whether a number can receive iMessage before anything sends
• send_message sends one message from a line you own
• get_conversation reads a thread's history with per-message delivery states
• check_message_status reports pending, sent or failed for one message

How do replies get back to the agent?#

Webhooks carry the inbound side: Blue Reacher posts every reply, delivery state and STOP opt-out to your HTTPS endpoint, signed, so your agent resumes the thread instead of polling.

• Delivery events are the useful half. Delivered but silent for two days is a follow-up trigger, and an RCS or SMS fallback tells the agent to write shorter, the way a person would on plain SMS.
• STOP is enforced platform-side before your code ever sees it, so compliance never depends on your prompt.

Where should a human stay in the loop?#

Blue Reacher treats send_message as the tool that touches a real person, and your agent should too: reading threads is safe to automate on day one, first-contact sends earn autonomy later.

• Rehearse on a brk_test_ sandbox key, where every tool answers with simulated data and nothing real sends
• Keep sends behind approval until the agent has a track record on replies
• Have the agent draft the batch, review it, then send approved messages one at a time
• Log the returned message id against the CRM record that triggered the send