Loading ToolBench
Runtime context MCP server for AI coding agents. Provides browser observation tools (console, network, screenshots) and action tools (click, type, navigate, run JS) to give AI agents eyes and hands in the browser.
Argus has 14 well-named tools with clear verb-noun patterns (get_*, list_*, click_*, type_*, etc.). All tools have descriptions (avg ~120 chars, within baseline 34-392). Input schemas are present with typed parameters and descriptions. However, output schemas are largely undocumented, most tools return JSON strings without formal schema declarations. Error handling is minimal; tools return plain text fallbacks ('No screenshots captured') rather than actionable recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Parameter constraints lack enums where appropriate (e.g., 'level' in get_console_logs accepts 'log'|'warn'|'info'|'debug'|'all' but is not declared as enum). Composition is strong, tools chain well (list_screenshots → get_screenshot, get_error_source_context uses error_index from get_console_errors). Security: no secrets in params, but no audit logging or permission gates visible.
Clear stored browser context.
Click an element on the page.
Get recent JavaScript errors from the browser console.
Get recent console output (log, warn, info, debug).
Parse a console error's stack trace into structured source locations. Returns the error message, parsed file:line:column for each stack frame, and identifies the primary app-code location (skipping node_modules, CDN, etc.). Use this to quickly find which source file to fix.
Get recent failed network requests (4xx, 5xx, network errors).
Output schemas undocumented. Most tools return JSON strings or TextContent without formal schema declarations. LLMs cannot reliably parse or chain results.
Parameter constraints not formalized as enums. 'level' in get_console_logs accepts 'log'|'warn'|'info'|'debug'|'all' but is a free-form string. 'event_type' in clear_context similarly lacks enum constraint.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Tools are classified as READ_ONLY or WRITE in metadata but not in MCP schema. Agents cannot infer safety without reading descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2026-07-28+ | v2 |
Get recent network requests (all, not just failures).
Get current browser page URL, title, and viewport size.
Get a browser screenshot by index. Use list_screenshots first to see the timeline, then fetch specific screenshots by index.
Get details about the UI element the user right-click captured. Returns selector, computed styles, HTML, bounding rect, and the page URL.
List all available screenshots with metadata (without image data). Returns a timeline of captured screenshots with index, URL, timestamp, trigger (hotkey/page_load/tab_switch/periodic/user_click), and viewport size. Use this to decide which screenshots to fetch with get_screenshot(index).
Navigate the browser to a URL.
Scroll the page to an element or position.
Type text into an input field.
Error handling lacks recovery guidance. Tools return plain text ('No screenshots captured') instead of actionable errors. No error classification (retryable vs user-fixable vs fatal).
No pagination support. Tools like get_console_logs, get_network_log accept 'limit' but no offset/cursor. Large result sets could exhaust context window.