Skip to main content

Web Access

Web Access v2 lets agents use a controlled browser when static fetching is not enough. It is useful for dynamic pages, signed-in dashboards, complex forms, single-page applications, internal systems, and pages that require screenshots.

Enabling

Web Access is a skill-scoped capability. Browser control, CDP proxy, site-pattern, and local-bookmarks tools are only exposed when the related skill is active. This prevents ordinary tasks from accidentally using browser tools and lets you know when an agent may interact with web pages.

Capabilities

CapabilityDescription
Tab controlOpen, list, switch, and close browser tabs
Page actionsClick, scroll, type, and select elements
ScreenshotsCapture page or element state for visual reasoning
CDP proxyUse Chrome DevTools Protocol for finer page state
File uploadSelect local files when needed; usually requires confirmation
Local bookmarksSearch Chrome/Edge bookmarks and history for URL hints
Site patternsRecord login entry points, selectors, action paths, and notes for a site

SitePattern

SitePattern stores per-site experience that helps the agent remember how to use a particular website:

  • Login entry points and common redirects
  • Locations of search boxes, filters, and export buttons
  • Common errors and popup-handling tips
  • Pages that require waiting for dynamic content

Validation is performed before saving a site pattern to prevent invalid or overly broad rules.

LocalBookmarks

LocalBookmarks searches the local browser's bookmarks and history for URL hints. It does not log in automatically or bypass permissions—it simply helps the agent find entry points to systems you use frequently.

Safety

  • Only http and https URLs are allowed by default
  • Actions with external impact, such as file uploads, form submissions, deletions, or publications, trigger approval
  • Browser sessions try to isolate tabs to avoid cross-task contamination
  • Screenshots and page content enter the current task context; avoid opening unrelated tasks on sensitive pages

Relationship to WebFetch / WebSearch

DesireCore's web capabilities are organized in three tiers. The agent automatically selects the lowest-cost option that can complete the task:

TierToolUse caseNotes
1WebSearchFind public information, news, and technical docsReturns search-result summaries without opening a browser
2WebFetchRead the body of a known URLSmart content extraction, ad/navigation removal, 15-minute cache
3Web AccessDynamic pages, signed-in dashboards, SPA forms, screenshotsLaunches a controlled browser and simulates real user interaction

If static fetching is sufficient, the agent does not open a browser.

Typical Scenarios

ScenarioRecommended toolReason
Look up API documentationWebSearchPublic information; a search suffices
Read a blog postWebFetchStatic page; direct content extraction
Check a management dashboardWeb AccessRequires login and page interaction
Operate a single-page applicationWeb AccessContent is dynamically rendered by JavaScript
Fill out and submit an online formWeb AccessRequires typing and clicking
Capture the current page stateWeb AccessRequires screenshot capability
Automatic Fallback

WebSearch itself has a fallback strategy: it prefers LLM server-side search, falls back to an independent search API, and only starts a browser sub-agent as a last resort. You do not need to choose manually; the agent decides based on the task.