Loading ToolBench
MCP server bridge for Agent Portal, providing tools for browser automation, agent steering, inspection, and report generation through the Agent Portal runtime.
This server has 33 tools with schemas and descriptions present, but quality is significantly below production baseline. Most tool descriptions are terse (10-30 chars), far below the 194-char average for A+ tools. Parameter descriptions are almost entirely absent, most tools have 0-1 parameter descriptions despite 1-5 parameters each. No output schemas are documented. Error handling is not visible in the tool definitions. The naming is generally verb-first (good), but many tools lack the specificity needed for LLM disambiguation (e.g., 'browser_open' vs 'navigate_to_url' both navigate; 'propose_action', 'approve_action', 'reject_action' form a workflow but lack dependency documentation). No tool has documented return types or pagination guidance. Security-sensitive tools like 'approve_action' and 'stop_agent' lack permission gate documentation.
Approve and optionally execute a queued action.
Navigate backward in history.
Close the current browser session.
Navigate forward in history.
Open a browser URL through Agent Portal.
Refresh the current page.
Return browser session status.
Parameter descriptions almost entirely missing. Of 33 tools with 1-5 parameters each, most have 0 parameter descriptions. LLMs cannot infer what 'selector', 'reason', 'mode', 'action_type' mean without explicit guidance.
Tool descriptions are uniformly terse (10-35 chars), far below the 194-char production baseline. Descriptions lack WHEN to use, prerequisites, or state-change implications. E.g., 'Refresh the current page' does not explain when to call it vs other navigation tools.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | 2025-06-18+ | v2 |
Capture a screenshot with evidence metadata.
Click an element selected by CSS, role text, or label.
Export a report to a destination path.
Generate a runtime report.
Get the runtime action queue.
Get the current runtime goal.
Hover over an element.
Inspect a specific element.
List available runtime reports.
Navigate to a URL through the runtime.
Pause the agent runtime.
Propose an action through the runtime policy engine.
Read the page accessibility tree.
Read collected console errors.
Read the current page DOM.
Read collected network failures.
Read visible text from an element.
Read a report by name.
Reject a queued action.
Resume the agent runtime.
Scroll the page or bring an element into view.
Set the runtime action mode.
Set the current runtime goal.
Stop the agent runtime.
Type text into an input element.
Wait for a page element to appear.
No output schemas documented. Tools like 'read_dom', 'read_accessibility_tree', 'generate_report' return complex structures but provide no schema guidance. LLMs cannot plan downstream tool calls without knowing what fields to expect.
Workflow tools (propose_action, approve_action, reject_action) lack dependency documentation. No description explains that approve_action requires an action_id from propose_action, or that these form a confirmation workflow. LLMs may misuse them.
Destructive/sensitive tools (stop_agent, pause_agent, approve_action, set_goal) lack permission gate or confirmation documentation. No description indicates these require authorization or have irreversible consequences.