Integrations
An integration connects SetForth to a tool you already use, so an agent can check its own work against your real production data instead of stopping at "the tests passed".
Once a project has an integration connected, new follow-up checks become available when you plan a task. A follow-up runs after the task's change ships. It reads the connected tool, and it passes or fails on what actually happened. A task with a failing follow-up is not done: the agent gets the result back and works the problem.
What you can connect
| Tool | What agents check |
|---|---|
| Sentry | New errors, error reduction, regressions |
| PostHog | Event tracking, adoption, funnel regressions |
| Amplitude | Event tracking, adoption, funnel regressions |
| Mixpanel | Event tracking, adoption, funnel regressions |
| Datadog | Latency budgets and new alerts |
| Grafana | Latency budgets and new alerts |
| New Relic | Latency budgets and new alerts |
| Railway | Deploy health after a release |
| LaunchDarkly | Flag rollout, kill switches, and stale flags |
Anything not on that list can still be connected as a custom connector, as long as it speaks the Model Context Protocol.
How to connect one
- Open your project, then Integrations.
- Pick the tool from the catalog.
- Authorize it, or paste the credential the tool gives you. Which one depends on the tool, and each page above says exactly which.
- SetForth tests the connection right away and tells you if it did not work.
Connecting is per project and needs the manage project permission. A credential you paste is encrypted and never shown again, and agents never receive it: SetForth makes the calls on their behalf and passes back only the result.
Read and write
Integrations start read-only. An agent can query the tool and nothing more.
Some tools can also change things, and a few follow-ups need that. Turning on write access on an integration's page opts into it for that project. Leave it off unless a check you want requires it.
Choosing follow-ups
Follow-ups appear during planning, once for each connected tool that can run them. If two tools cover the same ground, both show up and you pick. Each has its own settings, like how long to observe after the deploy, or how much of a drop in errors counts as a pass.
You can also switch individual follow-ups off for an integration, on its page under Skills. A follow-up that is off never appears during planning.
When something stops working
If a credential is revoked or expires, the integration's status changes and any follow-up waiting on it stops and asks for a person. Reconnect on the integration's page and the waiting work picks up again.
For connections authorized through a browser sign-in, the connection carries the access of whoever authorized it. If that person loses access to the tool, the integration stops working. Its page shows who authorized it, and Reconnect hands it to someone else.