Developers · MCP Beta

Let your assistant ask your cameras.

Add Watchful to Claude, ChatGPT or your own agent as an MCP connector. It can ask what your cameras saw, get alerts, and start or stop what Watchful watches for. Only with your consent, and only what you allow: answers and alerts, and at the third level pictures and short clips of the moments Watchful returns.

Quick start

Claude (claude.ai, the desktop app)

Settings → Connectors → Add custom connector. Name it Watchful, paste the address above, and press Connect. Claude opens Watchful's consent page: sign in, pick the stations and the access, Allow. Then ask: "Is anyone in the living room?"

Claude Code

claude mcp add --transport http watchful https://net.belike.vip/cloud/mcp

Then run /mcp in Claude Code and choose Authenticate for watchful.

ChatGPT and other assistants

Any client that speaks MCP over Streamable HTTP with the MCP authorization flow connects with the address alone; it discovers the rest. Where connectors are added in developer mode, add a custom connector with the address above and OAuth.

A self-hosted agent (OpenClaw and others)

No browser sign-in? The owner makes an agent key in their Watchful account and gives it to the agent, which sends it as a Bearer token:

curl -s https://net.belike.vip/cloud/mcp \
  -H "Authorization: Bearer wak_…" -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"ask","arguments":{"question":"Did anyone come to the door today?"}}}'

Use cases

The assistant brings the reasoning and its other tools (messages, a smart home, a calendar, a ticket system); Watchful brings the eyes. A few ways they work together:

Privacy and permissions

The short version: an assistant gets words about what the cameras saw, on the stations the owner picked, for as long as the owner allows. Pictures and short clips only at the third level, and only of the moments Watchful itself returns; it never changes how the station is set up, and the owner can see and stop it at any moment.

Answers say what they’re based on. Watchful tells the assistant which period it searched, and the assistant passes that on. It never claims to have watched the feed.

What a connection can and cannot do, by access

How it is enforced

• Twice: in the cloud and on the station. Watchful Cloud only offers and runs the tools the access allows. The station also gives the connection its own key, a device named "Agent · …", that reaches only a short list of requests for its level (questions, cameras, alerts; standing requests at Watch; pictures at See) and is refused everything else, whatever the cloud asked.
• Its own lane. An assistant's question never interrupts the owner's, it never answers a question the station put to the owner, and at Ask an instruction like "start watching" or "rename the camera" is politely refused.
• Seen by the owner. Everything it asks shows in the station's activity, marked with its name.

The owner's controls

What passes through, and what is kept

Sign-in (OAuth)

Watchful follows the MCP authorization spec: OAuth 2.1, authorization code with PKCE, public clients. A call without a token answers 401 with a WWW-Authenticate header that points to the metadata.

Redirect addresses must be https, localhost, or an app's own scheme, and must match what was registered. Codes are good once, for five minutes. Access tokens last an hour; refresh tokens rotate on every use and end with the connection (90 days, or the moment the owner turns it off). The granted access comes back as scope in the token answer.

Agent keys

For an agent that cannot open a browser for the owner, the owner makes a key in Connected agents → Make an agent key. It starts with wak_, is shown once, works as a Bearer token on the MCP address, has Ask access, and shows on its own card with the same Turn off, Renew and Pause all.

Alerts to agents

An always-on agent need not ask in a loop. Call subscribe_alerts with a public https address and Watchful POSTs each new alert there within seconds: falls, someone at the door, a standing request coming true, a camera going quiet. Alerts reach a webhook only while the connection is on and not paused, and only for the stations it was given. The owner sees on its card where alerts go and when the last one went.

POST https://your-agent.example/watchful
Content-Type: application/json
X-Watchful-Timestamp: 1792302173
X-Watchful-Signature: sha256=5d1c…

{"type":"watchful.alert","id":"3ec39f652826:1842","station":{"id":"3ec39f652826","name":"Dell"},
 "ts":1792302171.4,"time":"2026-10-18 09:02 UTC","camera":"outside","kind":"alert","severity":"",
 "headline":"It started to rain on the balcony.","detail":"At 9:02 AM.",
 "link":"https://www.watchfulbox.com/app/v/3ec39f652826?t=1792302165&cam=c85750bb","clip_available":true,"connection":"OpenClaw"}

Check every delivery before acting on it: the signature is an HMAC-SHA256 of the timestamp, a dot, and the raw body, with the secret subscribe_alerts gave you. Refuse old timestamps (say, over five minutes).

# Python
import hmac, hashlib, time
def genuine(secret, headers, body: bytes):
    ts = headers["X-Watchful-Timestamp"]
    want = "sha256=" + hmac.new(secret.encode(), ts.encode() + b"." + body, hashlib.sha256).hexdigest()
    return abs(time.time() - int(ts)) < 300 and hmac.compare_digest(want, headers["X-Watchful-Signature"])

// Node
const crypto = require("crypto");
const genuine = (secret, h, raw) => {
  const want = "sha256=" + crypto.createHmac("sha256", secret).update(h["x-watchful-timestamp"] + "." + raw).digest("hex");
  return Math.abs(Date.now() / 1000 - +h["x-watchful-timestamp"]) < 300 &&
    crypto.timingSafeEqual(Buffer.from(want), Buffer.from(h["x-watchful-signature"]));
};

Answer with any 2xx. A failed delivery is tried once more a few seconds later. Webhooks must be https on a public address.

Tools

tools/list shows only what the connection's access allows; a tool above it answers with an error that says so.

Answers are text, with a link to the moment and a "Watch live in Watchful" link (https://www.watchfulbox.com/app/live/<station>?cam=<camera>): live video is never sent to an assistant, it hands the owner that link. At see, an answer about moments carries one picture per moment, up to four (the latest); when there are more, it says how many and when, so the assistant can offer them and fetch each with get_picture and at. Clips are never attached: the answer says one is available, and get_clip makes it when the owner asks. Questions that work well: "Is anyone home?", "Did you see where I left my keys?", "When did the dog last go out?", "How many people came to the shop today?"

→ {"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"ask","arguments":{"question":"Is anyone in the living room?"}}}
← {"jsonrpc":"2.0","id":3,"result":{"content":[{"type":"text","text":"Nobody is in the living room.\nThe last person left at 11:33 PM.\n(Dell, 2026-10-17 21:24 UTC)"}],"isError":false}}

Protocol

MCP over Streamable HTTP. Send JSON-RPC with POST to the address; answers come back as application/json (no server-sent stream, so GET answers 405). Single messages and batches both work; notifications answer 202. Protocol versions 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05 are accepted. Methods: initialize, ping, tools/list, tools/call.

Errors

Coming next

• Listed connectors. Watchful in Claude's and ChatGPT's connector directories, so owners add it with one tap.

Building something with Watchful? Write to yogev77@gmail.com.