Web fetching
Let the agent read webpages it's pointed at — or lock it down completely.
Rivermark can fetch the contents and look of external websites — for example, to look up a URL provided in the task or references. The allow_fetch boolean controls this.
{
"task": "Create a carousel that explains the product on this page: https://acmecoffee.co/roaster. 5 slides.",
"allow_fetch": true
}- On
/design(defaultfalse): when enabled, the agent may fetch pages it's explicitly pointed at in the task or references. - On
/brandkit(defaultfalse): when enabled, the agent may follow links beyond the provided URLs — reaching an "About" page, a blog index, or a linked media kit zip — to build a more complete picture of the brand.
When allow_fetch is false, a URL reference still works: the page is fetched once, extracted, and used as reference material (see References). The flag only governs the agent following additional links at runtime.
Sandboxing
Fetches happen inside the agent's sandbox, per-request. The platform prevents agents from accessing other jobs' data; fetching is additionally scoped to public web content only.
When to enable it
Enable allow_fetch when your task points at URLs whose full context matters — "redesign this page", "make a summary of this blog post", "extract the brand from this site". Keep it off for mechanical tasks, where deterministic inputs keep costs and runtimes predictable.