One-Tap Approve/Reject for AI Agent Actions

Give a human a single tap to approve or reject what an AI agent wants to do — from a phone notification, no dashboard tab required, in seconds.

The scenario: a support-reply agent that can't be trusted alone

Say you have an agent that reads incoming support tickets and drafts replies. Most drafts are fine. Some aren't — a misread ticket, a wrong refund amount, a tone that doesn't fit an angry customer. You don't want to review every reply in a web dashboard between other work; you want a push notification that shows the draft and lets you tap Approve or Reject and move on.

That's the entire feature Impri is built around: the agent pushes a proposed action, a human gets notified wherever they already look (phone lock screen, Telegram, Slack), and one tap resolves it. No typing, no dashboard, unless something actually needs a closer look.

The flow in code

Three calls. The agent doesn't need to know which channel the human taps from — that's a notification setting, not an API parameter.

import os, time, requests

API = "https://api.impri.dev"
HEADERS = {"Authorization": f"Bearer {os.environ['IMPRI_API_KEY']}"}

def request_approval(ticket_id: str, reply_body: str) -> dict:
    resp = requests.post(f"{API}/v1/actions", headers=HEADERS, json={
        "kind": "support.reply",
        "title": f"Reply to ticket #{ticket_id}",
        "preview": {"format": "markdown", "body": reply_body},
        "target_url": f"https://helpdesk.example.com/tickets/{ticket_id}",
        "expires_in": 3600,  # stale after an hour, don't send a late reply
    })
    return resp.json()

def wait_for_decision(action_id: str) -> str:
    while True:
        r = requests.get(f"{API}/v1/actions/{action_id}", headers=HEADERS).json()
        if r["status"] != "pending":
            return r["status"], r
        time.sleep(5)

action = request_approval("48213", "Hi, thanks for reaching out — I've issued a full refund...")
status, result = wait_for_decision(action["id"])

if status == "approved":
    send_reply(result["decision"]["final_preview"]["body"])  # your send function
    requests.post(f"{API}/v1/actions/{action['id']}/result", headers=HEADERS,
                   json={"status": "executed"})

Note there's no editable field here — this action is deliberately approve-or-reject only. If you also want the human to be able to tweak the wording before it goes out, see editing drafts before approving; that's a different, slightly heavier flow.

Where the tap actually happens

The POST /v1/actions call is channel-agnostic. Where the human taps depends on how notifications are configured for their account, not on anything the agent sends:

See notifications for setup, or Telegram approval and Slack approval for the two most common one-tap setups.

What "one tap" does and doesn't guarantee

The tap resolves status to approved or rejected — that's the whole contract. It does not mean Impri checked the content is correct; it means a specific human looked at the title and preview and made a call. The gate only holds if the agent has no other way to send the reply — if it also holds a raw API key to the helpdesk and can call it directly, a human tapping "reject" here doesn't stop that path. Wrap the actual send function so the approved decision is the only way in; see the MCP server for a pattern that does this without writing the polling loop yourself.

Rejected and expired look the same to your code

Both leave status as anything other than approved, and in both cases the execution branch is simply never reached. A ticket reply that expired after an hour is exactly as unsendable as one someone explicitly rejected — treat them identically rather than special-casing expired.

Next: quickstart to get an API key and push your first action, or MCP if your agent runs inside Claude Code or Claude Desktop.