# Connect your AI

Let your own AI ask you things: a quick yes or no, a choice, a link to open, or a step only you can do. Each ask waits for you in
the person.tools app or on your person.tools page, under the name you gave that AI; one your AI marks "now" also sends a
notification to your phone. You answer with a tap, and your AI reads the answer.

What your AIs send you is yours. Nobody at person.tools looks at it. Photos and clips are stored encrypted and deleted after
seven days, or when you choose. Nothing you get here is ever shown to a reviewer.

## 1. Make the connection (you)

On your person.tools page, go to **Your AIs** (in the person.tools app: **You**, then **Connect an AI**). Give it a name (for example "My
assistant" or "Home camera") and an icon. You get a **key** (it starts `conn_`), a **link**, and a **webhook secret** (shown once). The key does exactly one thing:
it lets that AI ask you, and read your answers. You can revoke it with one tap; you can also quiet it or block it.

No person.tools page yet? Make your own at https://run.person.tools/start: a minute, no email.

## 2. Give your AI the key

**If your AI can add an MCP server with a header** (Claude Code, Gemini CLI, Cursor and others): add
`https://api.person.tools/mcp` with the header `Authorization: Bearer YOUR_KEY_HERE`. It sees exactly two tools: `ask_person`
(title, step, unlocks, urgency) and `ask_person_reply` (task_id). For Claude Code, one line:

```
claude mcp add --transport http ask-me https://api.person.tools/mcp --header "Authorization: Bearer YOUR_KEY_HERE"
```

**If your AI can make web requests** (an agent, a program, an assistant with a web-request tool), give it the key and this
prompt (copy it, put your key where it says). A chat that can do neither can't send an ask yet.

```
You can ask me things through person.tools when you need my decision or an answer only I can give.

To ask: POST https://api.person.tools/v1/tasks
Headers: Authorization: Bearer YOUR_KEY_HERE and Content-Type: application/json
Body: {"title": "a few words", "step": {"type": "question", "text": "your question", "choices": ["Yes", "No"]},
       "unlocks": "what my answer lets you do", "urgency": "today"}
It answers with a task_id.

To read my answer: GET https://api.person.tools/v1/tasks/TASK_ID with the same Authorization header. When "state" is
"done" or "answered", "reply" holds my answer: "reply.choice" is the choice I picked, "reply.text" a line I wrote, and
"reply.action" says how I answered (answer, done, approve, deny or not_now). Check every minute or so; I answer when I see it.
Or add "webhook": "https://your-address" to the ask: my answer is POSTed there when I give it, signed. The header
x-person-tools-signature is "v1=" and the HMAC-SHA256 of "<x-person-tools-timestamp>.<body>" with this webhook secret:
YOUR_WEBHOOK_SECRET_HERE. It is retried for up to 24 hours.

Rules: one ask, one step. A question has 2 to 4 choices, or leave out "choices" and I'll reply Done, Not now, or a line.
Other steps: {"type": "approve", "text": "exactly what you will do"} (I approve or deny; "reply.approved_text" is what I
approved), {"type": "link", "url": "https://...", "text": "what to do there"}, or {"type": "command", "command": "...",
"where": "where I run it"}. Urgency: "now" (first in my list, and one notification to my phone outside my quiet hours;
the ask's "rang" is true once it went out), "today" or "whenever" (in my list, no notification). To retry
safely, add "idempotency_key" (your own id for the ask, up to 128 characters): the same one again returns the first ask,
never a second. The title is 3 to 120 characters, "unlocks" up to 300, the step's text up to 4,000. For a decision that only matters for a while, add "expires_in"
(seconds, 60 to 2592000; 30 days if you leave it out): when it passes unanswered, the ask lapses and you're told (state
"lapsed"); it is never answered for me. Never put a password, a code or a key in an ask.
```

If your AI can only call an address, it can POST the same body to your connection's **link** instead (no header), and read
the answer at the link followed by `/TASK_ID`.

## 3. What happens next

Your AI places one ask. It appears in your list under the AI's name, on your person.tools page or in the person.tools iPhone
app (in testing now). An ask your AI marks "now" also sends one notification to your phone, outside your quiet hours (9 PM to 8 AM
unless you change them), at most one every 30 minutes, and not from an AI you quieted, whether your calls are
On or Off. On your page, that needs the page added to your Home Screen with notifications allowed. It plays a sound unless you turned sounds off (in
the app: You, then Sounds). Asks marked "today" or "whenever" wait in your list without a notification. You tap your answer, and
your AI reads it from the same address. If you'd rather not be asked, revoke the connection: its key stops at once and its open asks are withdrawn.

Asking you is free: no money, test or real. If you don't answer, the ask lapses after 30 days, or sooner if your AI set a shorter
time (from one minute up), and your AI is told; it can also withdraw the ask. There is no default answer, so your AI decides what to do while it waits (a door stays
locked, a refund waits).

## How this differs from a notification app

ntfy, Telegram or Home Assistant are free, older and fine for an alert. What person.tools adds: your answer goes back to your AI
as data it can act on (the choice you picked, the exact text you approved, and when); all your AIs reach you in one place, each
under its own name, and any one can be revoked, quieted or blocked with one tap; you set when you can be asked; and nothing an AI
sends you is read by anyone at person.tools.

## What "preview" means

what works now is real. Your own AIs ask you, and your own people answer your calls, for free. Strangers are not offered yet, so any money you see is test money only: nothing is charged and nobody is paid until a stranger pool opens. Until then, a stand-in for strangers is open to any test key (its pool is named in the discovery file, under strangers_stand_in): scripted stand-ins answer in seconds, every answer is marked simulated, and nobody at the pool is asked or paid. person.tools is run by Hammer Corp., Austin, Texas.

person.tools is run by Hammer Corp., 1401 Lavaca St #41100, Austin, TX 78701. Questions: hello@person.tools. What we keep, and
for how long: [/privacy](/privacy).
