Create a project

A project groups related agents, integrations, and variables - and decides who can reach them.

1. Go to Projects and choose Create Project.
2. Name it after the work, e.g. "engineering digests".
3. Set its visibility: Organization so your team can use it, or Restricted to keep it to yourself.

You already have a Personal project if you'd rather just experiment.

Connect a tool

1. Open Integrations and pick a service - GitHub is a good first one.
2. Choose an authentication method. The Helios app is the fastest; you can also bring an API key or your own OAuth app. See authentication.
3. Authorize, and give the integration a name you'll recognize.

Your agents can now call any endpoint of that service's API. There's no action list to pick from.

Ask an agent to do it

Open Chat and describe the task as you would to a new colleague:

Look at the pull requests opened in acme/api over the last week.
Summarize what shipped, and call out anything touching authentication.

Watch what the agent does - you'll see it search the GitHub API, make calls, and reason about what came back. If the result isn't what you wanted, say so and let it correct.

Put the agent on the task permanently

Once the instructions work, they're a workflow.

1. In your project, choose Create Workflow and name it "Weekly PR digest".
2. Paste the prompt. Add whatever the unattended version needs - where the result should go, and what to do when there's nothing to report:

Every Monday, look at the pull requests opened in acme/api over the last
week. Summarize what shipped and call out anything touching authentication.
Post the summary to #eng-standup in Slack.
If no pull requests were opened, do nothing.

3. Add a cron trigger. Type every Monday at 9am Eastern and Helios converts it to a cron expression for you, in UTC.
4. Connect Slack too, since the prompt now needs it.

Run it and watch

Dispatch the workflow manually rather than waiting for Monday. Every execution is a run with the agent's full transcript: every tool call, every response, and the final output.

If the agent did the wrong thing, the transcript shows you exactly where it went astray. Adjust the prompt and run it again.