Building your own DIY agent for incident resolution?

→

Understand what to look out for.

Building your own DIY agent for incident resolution?

→

Understand what to look out for.

Building your own DIY agent for incident resolution?

→

Understand what to look out for.

Triggers

One URL. Every system.

Herald is Hyground's external-notification gateway. A single endpoint accepts any JSON over HTTP, and an LLM chain decides whether it is an alert worth an RCA or just context to summarise. Alertmanager, Grafana, PagerDuty, Datadog, GitHub, Jira, ServiceNow, cron, curl: anything that speaks JSON wires up the moment it can POST.

Triggers

One URL. Every system.

Herald is Hyground's external-notification gateway. A single endpoint accepts any JSON over HTTP, and an LLM chain decides whether it is an alert worth an RCA or just context to summarise. Alertmanager, Grafana, PagerDuty, Datadog, GitHub, Jira, ServiceNow, cron, curl: anything that speaks JSON wires up the moment it can POST.

Triggers

One URL. Every system.

Herald is Hyground's external-notification gateway. A single endpoint accepts any JSON over HTTP, and an LLM chain decides whether it is an alert worth an RCA or just context to summarise. Alertmanager, Grafana, PagerDuty, Datadog, GitHub, Jira, ServiceNow, cron, curl: anything that speaks JSON wires up the moment it can POST.

Endless integrations

Integration is the URL, not code

No per-source adapters. Two LLMs (an injection_judge and an alert_extractor) reason over payload semantics, so any new producer is wired up the moment it can POST JSON to /api/herald/webhook. Webhooks, cron jobs, scripts, monitoring stacks, ticket systems, deployment tooling all enter through the same door.

Alertmanager, Grafana, Datadog, PagerDuty

Forward existing alert payloads as-is. No template authoring, no per-receiver mapping. Herald reads them and decides whether they warrant an RCA.

Title

Issue updates, deployment events, change tickets. Anything you would otherwise wire into a notification rule can become a Hyground investigation.

Cron, curl, anything that speaks JSON

A cron job that posts a daily report. A bash one-liner from your laptop. A custom script in a pipeline. If it can POST JSON, it can be a trigger.

How Hyground protects the webhook endpoint from injection

Every payload needs to pass both gates before a single agent token is spent on it.

Who can call the endpoint?

oauth2-proxy, mTLS, shared-secret, or IP allow-list, picked per deployment. The same gate the rest of Hyground uses, applied at the webhook ingress.

Who can call the endpoint?

oauth2-proxy, mTLS, shared-secret, or IP allow-list, picked per deployment. The same gate the rest of Hyground uses, applied at the webhook ingress.

What reaches the agent?

50 KB payload cap, JSON-shape check, an injection_judge LLM, and `<untrusted_content>` -fenced framing. Payloads scoring above the injection threshold are rejected before they ever reach the agent. The framing template tells the model to treat anything inside the tag as data, never instructions.

What reaches the agent?

50 KB payload cap, JSON-shape check, an injection_judge LLM, and `<untrusted_content>` -fenced framing. Payloads scoring above the injection threshold are rejected before they ever reach the agent. The framing template tells the model to treat anything inside the tag as data, never instructions.

Does a trigger call block while the investigation runs?

Webhook callers get a fast acknowledgement, but investigation happens in the background.

01

Producer posts JSON

Herald returns 202 Accepted in milliseconds. The caller is done: no long-lived connection, no retries to babysit.

01

Producer posts JSON

Herald returns 202 Accepted in milliseconds. The caller is done: no long-lived connection, no retries to babysit.

02

Triage + RCA in the background

The injection_judge classifies risk and shape. The alert_extractor structures alert-like payloads. The dispatcher opens a session on the root agent with the framed payload as the first message.

02

Triage + RCA in the background

The injection_judge classifies risk and shape. The alert_extractor structures alert-like payloads. The dispatcher opens a session on the root agent with the framed payload as the first message.

03

Notify Slack or Teams

An alert-received card lands immediately so the team knows Hyground is on it. When the session completes, a second card links to the full transcript and findings.

03

Notify Slack or Teams

An alert-received card lands immediately so the team knows Hyground is on it. When the session completes, a second card links to the full transcript and findings.

Want to go deeper?

Triggers compose with the rest of the platform

An incoming trigger is just an entry point. Once a session is open, every Skill, every connected adapter, every scheduled workflow is available to it.

Skills

Reusable operational procedures the agent reaches for when a trigger arrives. Defined once, executed consistently.

Explore Skills

➞

Scheduling

When no external system fires, the clock can. Same agent, same Skills, same investigation surface, on a cadence you set.

Explore Scheduling

➞

How the trigger endpoint is authenticated and governed

Herald is the front door. The same governance that applies to any Hyground session applies here: read-only by default, every investigation auditable, every action recommended with evidence to a human in the driver's seat.

The ingress gate decides *who* can post. The semantic gate decides *what* reaches an agent. The session log captures *everything* that happens after.

See Hyground in action

Book a demo

Try our sandbox