> ## Documentation Index
> Fetch the complete documentation index at: https://docs.use.observer/llms.txt
> Use this file to discover all available pages before exploring further.

# Probes vs scraping

> Why the agent runs probes from inside your network instead of having the cloud scrape endpoints.

The classical observability pattern (Prometheus, Datadog,
Grafana Cloud) is centralised scraping: a central system reaches
out to your endpoints on a schedule and pulls metrics. Observer
takes the inverse position: the agent runs in your network and
pushes outcomes to the cloud. This page covers why.

## The constraints scraping puts on you

Scraping requires the central system to reach every endpoint it
measures. In practice this means at least one of:

1. Public exposure of internal endpoints, sometimes with a
   reverse proxy or TLS-terminating load balancer purely to
   accept the scrape.
2. A VPN or peering connection from the central system back into
   your network.
3. A scraping agent inside your network that the central system
   pulls from (essentially shifting the same problem one hop).

Each option grows the attack surface, and in many organizations
opening an inbound path needs a security review.

## The probe model

Observer's agent runs inside your network, hits its targets
locally, and pushes each value with its verdict to the cloud. No IP allowlist at
your edge, no TLS-terminating proxy in front of internal
endpoints, no reverse VPN. The exact request surface (which
endpoints, which payloads, what stays put) is enumerated in
[Agent and cloud boundary](https://docs.use.observer/agent/concepts/agent-cloud-boundary.md).

The trade-off: collection happens in your network, so collection
runs on your hardware. The agent is a single process and runs
anywhere a container can run.

## When scraping is still preferable

When the targets are themselves SaaS systems with public scrape
endpoints (a third-party API, a public DNS server, a hosted
queue), the central system would have a clean path. Observer
still uses the agent for these cases for one reason: a single
configuration surface. Mixing "some metrics scraped centrally,
some pushed from an agent" doubles the operator's mental model
without an offsetting benefit. The agent runs the probe from
wherever it sits and reports the result.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.