Growth Method
Get startedSearch
Documentation menu

How approvals work

The AI reads from your connected tools freely and asks you before it changes anything. Here is how Growth Method tells the two apart, so you know when to expect an approval card.

What runs on its own and what asks

Every tool an integration offers can carry safety hints from its vendor: whether it only reads, and whether it can delete or overwrite something. Growth Method treats those hints as a starting point, not a promise, and applies three rules:

  1. Tools marked read-only run on their own. Looking up a report, listing pages or running a query happens without interrupting you.
  2. Tools marked as able to delete or overwrite always ask. You see an approval card every time, whatever the request.
  3. Everything else is checked request by request. Many tools say nothing either way, and some can both read and change things depending on what they’re asked. Before one of these runs, an independent check reads the request. If it’s confident the request only reads, it runs on its own. Otherwise you get an approval card.

You answer a card with Approve or Reject. See Chat with the AI assistant for what happens next.

Why we built the check

Some popular integrations, PostHog among them, offer a single tool that both reads and changes things, so before the check nearly every request they made showed a card, almost all of them plain reads. When most cards are noise, it’s easy to start approving without looking.

The check is one safeguard among several. See the Security overview and our security details for the rest.

The independent check

The check is a separate AI model, Jev from TypeSafe, reached through OpenRouter. It answers one question: would this request change anything?

  • It sees only the request: the tool’s name, the vendor’s description of it, and what the AI asked it to do. It never sees your integration’s keys or sign-ins, and neither provider trains models on what it sees.
  • When in doubt, it asks you. If the check is at least 80% confident a request only reads, it runs without a card. Anything less and you get an approval card, so a mistake costs you an extra click, not an unwanted change.
  • It judges each request, not the whole tool. The same PostHog tool can run a query on its own and still ask before creating a feature flag.
  • It’s quick: the check adds about a third of a second before a request runs.
  • If the check fails, takes too long, or the request is too large to read, you get an approval card.
  • A cleared request looks like any other read. It shows in the chat as a normal step, with no card and no extra marker.

Examples

PostHog shows the check at work, because one PostHog tool, exec, does everything: reading data and changing it go through the same tool. These two requests went to that one tool. The first is a real request from Growth Method’s own workspace; the second is from the set of reads and writes we test the check against before we rely on it. Here is exactly what the check saw for each.

A query that only reads. The AI asked for pageview and newsletter sign-up counts on one article (shortened here):

Tool: exec (PostHog)
command: call execute-sql {"query": "SELECT event, count() AS events FROM events
  WHERE event IN ('newsletter_subscribed', '$pageview')
  AND properties.$current_url LIKE '%forward-deployed-engineer%' GROUP BY event"}
context: Counting pageviews and newsletter sign-up events on a single article page.

It ran without a card.

A request that changes something. The AI wants to create a feature flag that shows a new pricing page to 10% of visitors:

Tool: exec (PostHog)
command: call create-feature-flag {"key": "new-pricing-page", "name": "New pricing page",
  "filters": {"groups": [{"rollout_percentage": 10}]}}
context: Creating a feature flag to roll the new pricing page out to 10% of visitors.

You get an approval card, and nothing changes in PostHog until you approve it.

Agents that run on a schedule

The same rules apply when a custom agent works on its own:

  • Human in the loop agents pause when a step needs approval and email your team a link to approve it.
  • Autonomous agents, once you’ve accepted the responsibility, don’t stop for approvals on their own runs. Any change they make in a connected tool without review is marked Ran without review in the chat. Lookups that only read data aren’t marked.

In a chat, the rules above apply whichever agent you’re working with.

Claude, ChatGPT and other AI apps

An AI app connected to Growth Method’s MCP server follows the same rules when it uses your integrations, once a manager has switched this on. Reads and requests the check clears run straight away. Anything else waits for you:

  • In AI apps that can show interactive cards, such as Claude, the approval card appears right in the chat. Choose Approve or Reject, and the AI carries on.
  • In other apps, you get a link to the same card in Growth Method. Open it, choose, then tell the AI. Select Show the full request to see exactly what the AI will send, however detailed.

Only the person who asked can approve or reject a request, and it expires after 24 hours. The AI is never offered the Approve button: in apps that show the card, it’s there for you to press, and everywhere else the link only opens for you, signed in to Growth Method. Each approval runs the change once.

Next steps