Loading ToolBench
Portable website capabilities, sessions, and repositories for any MCP agent
AnyBridge provides 39 tools with complete input schemas and descriptions for all tools. Naming follows verb_noun conventions (read_page, navigate, click_ref). Descriptions are substantive (avg ~120 chars) and explain WHEN to use each tool. However, parameter descriptions are inconsistent, some params lack detail on constraints, ranges, or expected formats. Output schemas are not documented in the tool definitions. Error handling guidance is minimal. Security considerations (credentials, permissions) are not visible in tool definitions. The server demonstrates solid foundational quality but lacks the polish and completeness expected of production-grade tools.
Call a WebMCP tool registered by the current website by name. Prefer the site's direct MCP tool when it is visible; this is the compatibility fallback.
Click a link or button by its visible text (or a CSS selector) and return the resulting page. Useful for navigation, cookie banners, and 'load more' buttons.
Click an exact element ref from the latest snapshot and return fresh page state.
Show the current browser URL, page title, and number of discovered WebMCP tools.
Show the adaptive engines available in this AnyBridge installation.
Extract structured page data using a JSON object whose values are CSS selectors, {selector, attr} definitions, nested fields, or one-item arrays for repeated rows.
Output schemas not documented. Tool definitions lack return type specifications, forcing LLMs to infer structure and risking misuse of returned data.
Parameter constraints under-specified. Numeric params (max_chars, limit, timeout_ms) lack min/max bounds; string params lack format/pattern guidance. LLMs may pass invalid values.
Error handling guidance absent. Tool descriptions do not explain recovery paths (e.g., 'If navigation fails, try reset_session'). LLMs lack actionable error context.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Fill an exact input ref without exposing the value in any saved workflow.
List the forms on the current page with their fields, so submit_form can be called.
List links on the current page (text and URL), optionally filtered by a substring.
List saved browser profiles (cookies, local storage, etc.).
List Git repository aliases saved in AnyBridge.
List website aliases saved in AnyBridge. Use this when the user refers to a site by a familiar name instead of giving its URL.
List the WebMCP tools registered by the current website. Use this after navigation if newly discovered site tools are not already visible to you.
List saved browser automation workflows.
Open a URL in the browser and return the page content as markdown.
Watch for changes on the page and report them as they occur.
Load a saved browser profile (cookies, local storage, etc.).
Clone or reopen a Git repository and return its absolute path.
Open a saved Git repository alias and return its absolute path.
Open a saved website alias in the browser.
Press a keyboard key, optionally on an element ref.
Read the current page as markdown. Use a CSS selector to scope to one part of the page, and max_chars to read more of a long page. If the browser is showing a PDF, returns its extracted text with page markers; use pages="12-15" to read a specific range of a long document.
Remove a saved browser profile.
Remove a saved Git repository alias (local files are not deleted).
Remove a saved website alias.
Remove a saved browser automation workflow.
Recover the AnyBridge browser after a crash or broken page.
Execute a saved browser automation workflow.
Save the current browser session as a named profile.
Save a Git repository URL under a memorable alias.
Save a website URL under a memorable alias for quick access.
Save a sequence of browser actions as a named workflow.
Capture the current page as PNG when semantic content is insufficient.
Choose a value in an exact select-element ref.
Read a URL through AnyBridge's continuity engine. It uses durable cache, HTTP, Lightpanda or Chromium, and Wayback as a historical last resort. For PDF documents it returns extracted, page-marked text (never screenshots): use pages="12-15" to read a specific range of a long document.
Return a compact semantic snapshot. Interactive elements receive refs such as e3; prefer ref tools over text or CSS selectors.
Fill and submit a form by its index from list_forms; returns the resulting page as markdown.
Type into an input found by placeholder, label, or CSS selector — works even for inputs outside a <form> (single-page apps). Optionally press Enter to submit.
Wait for visible text or a CSS selector, then return a fresh snapshot.
Destructive operations lack confirmation pattern. Tools like run_workflow, remove_workflow, and remove_saved_site execute irreversibly without dry-run or confirmation step.
Security and permissions not declared. Tool definitions do not specify required scopes, permission checks, or audit trail requirements. No guidance on secret handling.