Settings
The defaults are honest. Everything here is for going further.
Fivetran & destination
Bring your own Fivetran account
By default Graft deploys into its own account. Paste your Fivetran API key + secret to graft straight into yours — the connector outlives this site.
Bring your own BigQuery
Land rows in your project instead. Paste a service-account JSON with dataset-create rights.
Dataset retention
House datasets are swept on schedule. Your own BigQuery is never touched by the sweeper.
Sync & backfill
Sync schedule
Fivetran's scheduler, not a cron we wrote.
Backfill cap
Historical rows pulled on the first sync. A cap is a courtesy to the tool on the other end.
Usually auto-detected. Set it only if the agent guesses wrong.
Agent behavior
Strategy
docs first
probe first
Probe-first skips the docs even when given — useful when the docs lie.
Pagination hint
cursor / offset / page number / none — set it only when auto-detection fails.
Auth style
How the key is presented to your tool. Custom header available.
Patch attempts
3
How many times the agent may read failing logs and rewrite its own connector before giving up honestly.
Review code before deploy
Pause after the sandbox test. Nothing deploys until you've read connector.py and pressed approve.
Privacy & data
Redact sample values in the feed
Sandbox tables show shapes, not contents, until you flip this.
Developer
Show raw MCP traffic
Every tool call the agent makes to Fivetran, inline in the feed.