Loading ToolBench
Claude Code 的全交互对外抓手 MCP —— 搜/抓/登录态抓/控桌面，四通道+多引擎 fallback+AES cookie+markdown 抽取。All-in-one interaction MCP for Claude Code.
Lasso has 21 tools with mixed quality. Most tools have basic descriptions (50-150 chars) and minimal parameter documentation. Input schemas are present but sparse: many tools accept only 1-2 params with minimal type info. Critical gaps: no output schemas documented anywhere, no error recovery guidance, no parameter constraints (enums, ranges), and descriptions lack WHEN/WHY context. Tool names are action-oriented (search, browse, fetch) but several are vague (desktop, admin, interact_act). The server exposes destructive operations (admin, download cancel, desktop control) without permission gates or confirmation patterns. Composition is weak, tools like browse_headless, browse_logged_in, browse_headed, and browserbase/steel all do similar things with no clear differentiation. No pagination support visible. Overall: functional but well below production baseline.
Administrative actions for runtime capability management
Browse web pages in headed mode with user interaction
Browse web pages in headless mode
Browse web pages with login state
Browse using Browserbase cloud browser service
Control desktop via accessibility APIs, AppleScript, or CGEvent
Run diagnostic checks on Lasso server health and configuration
No output schemas documented. Tools return unstructured responses (e.g., screenshot, pdf, network, doctor, admin) with no field definitions. LLMs cannot plan downstream calls or extract structured data.
Destructive/sensitive tools (admin, desktop, download cancel, interact_act) lack permission gates, confirmation patterns, or dry-run support. No audit trail or user authorization checks visible.
Parameter descriptions are minimal or missing. 'action' params (desktop, interact_act, admin) have no enum constraints or format guidance. LLMs must guess valid values.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | <=2025-11-25 | v2 |
Download files with start/status/wait/cancel actions
Fetch and parse RSS/Atom feeds
Fetch and extract content from a URL with multiple extraction modes
Perform actions on interaction forest elements
Observe current state of interaction forest
Query interaction root elements for forest-based desktop automation
List network requests from current browser session
Generate PDF from current browser page
Read text content from spilled files with offset support
Take a screenshot of the current browser state
Search across multiple sources (Brave, machine MCP, headless browser)
Search local data sources (Chrome History, Spotlight files)
Browse using Steel cloud browser service
Look up archived versions of URLs via Wayback Machine
Four browse tools (browse_headless, browse_logged_in, browse_headed, browserbase, steel) perform similar operations with no clear differentiation. Composition violates single-responsibility principle.
No error recovery guidance. Tools lack descriptions of failure modes, retryability, or next steps. Error responses will not guide LLM behavior.