Sign inGet started
Skills

Skills security

View as MarkdownOpen in ChatGPTOpen in Claude
Connect your agent
Add to CursorAdd to VS Code

What a scheduled analytics run can read and do, whose Claude Code runs it, what it can never touch and the limits that keep your website safe.

Updated

A scheduled run reads your project's data and reports what it found. It never changes your website or any tool. The rest of this page is the detail behind those two sentences.

Whose Claude runs a skill?

Runs happen in the Claude Code of the person who scheduled the skill, on their plan, with their own access. A teammate's skill never runs on your Claude Code. We run no model and never use your data to train one.

What can a run do?

A run gets one ticket from start_run, which names the tools it uses: reading tools such as get_dashboard, query and find_changes, and post_finding, suggest_memory and finish_run. It works on one project, the one its skill belongs to. It runs with your own access, so it can never reach more than you can: a project you aren't on, or an action your role doesn't allow, is refused. If someone edits the skill after the run was scheduled, the run stops.

What does a run never do?

The run's instructions rule these out:

  • Change your website, your code or another tool.
  • Act through a connector.
  • Read another project.

Who makes the changes?

People do. When a finding proposes a change, a person asks their agent to make it: in their code with their approval, or through a connector an owner or admin added. Every connector call is logged with who made it, and any connector can be blocked for the team.

Next step

How a run works

Related
  • Run skills on a schedule: Allow scheduled skills on your Claude Code once: a cloud routine with a token, or a one-line hourly check.
  • Skills: Skills are instructions your agent follows on demand or on a schedule. They tell you what changed and what to fix, and stay quiet otherwise.
  • How a run works: A scheduled run, step by step: it starts, reads context and numbers, reports a finding or stays quiet and finishes.
  • Anatomy of a finding: One finding from start to fix, told as a story: our rule notices wasted ad spend, a skill runs and explains it, and a person decides what to change.
  • Project knowledge: The team's instructions, goals and memory your agent reads before every answer and every run, who writes them, the limits and how agents add to them.
  • Steer skills: Shape what skills look at and how they report, with team instructions, a skill's own text and feedback.
Still have questions?
Was this page useful?