# The Backwatch bot

If you found this URL in your logs, someone ran Backwatch against your site. Backwatch checks how AI agents see a site: whether they can read it, find its MCP server, sign in and find it in agent registries.

## How to recognise it

```text
Mozilla/5.0 (compatible; Backwatch/0.1; +https://backwatch.dev/bot)
```

Backwatch always sends this User-Agent. It never retries in disguise, even when it is blocked. If your site blocks it, the report says so. A site that blocks automated agents is something its owner usually wants to know.

The AI-bot checks add the crawler's name, so your robots.txt and firewall rules for bots such as GPTBot and ClaudeBot are tested, while the request still says it is Backwatch:

```text
Mozilla/5.0 (compatible; GPTBot/1.0; via Backwatch/0.1; +https://backwatch.dev/bot)
```

Those requests come from the same machine as the rest of the scan, not from the AI companies.

The site owner can also choose a different User-Agent with `--user-agent` when testing their own site.

## What it is not

Backwatch is not a crawler. It does not follow links across the web, build an index or train models. It runs only when a person starts a scan, from their own machine or their own CI, so there are no fixed IP addresses.

## What it requests

A scan is small and bounded: at most 100 requests to at most 12 hosts, 10 content pages and 3 minutes. Typically:

- the home page and a few linked pages, also with `Accept: text/markdown`
- `robots.txt`, `sitemap.xml`, `llms.txt` and files linked from them
- `/.well-known/` discovery files (MCP, OAuth, API catalogues)
- published API descriptions such as OpenAPI documents
- one request to a path that does not exist, to see the 404 an agent gets
- MCP and API endpoints without credentials, to read their sign-in challenge

It sends no credentials, submits no forms and makes no requests that change data. It does not register OAuth clients unless the owner of the server passes `--register` for their own server.

Backwatch reads `robots.txt` to report on it. Because a person starts each scan on demand, it does not obey `Disallow` rules, the same way other user-triggered fetchers work.

## To block it

Match `Backwatch` in the User-Agent and return 403. The person who ran the scan will see "blocks automated agents" in their report.

## Contact

To be added before launch.