Skill + browser MCP for coding agents behind a full-tunnel VPN: the app under test reaches domestic services over the real ISP while the agent stays on the tunnel.
This MCP server has severe definition quality issues. Tool definitions are inferred from CLI argument parsing rather than explicit MCP tool registration with schemas. No tool has a visible input schema in JSON Schema format, parameters are documented only in descriptions as free-form strings. Descriptions exist but are brief (50-120 chars) and lack actionable guidance for LLM selection. No output schemas are documented. The 'proxy' and 'install-skill' tools perform destructive operations (WRITE risk) but lack confirmation patterns or error recovery guidance. Tool names are generic verbs ('browser', 'proxy', 'detect', 'route') that do not clearly convey intent or distinguish from similar tools. Parameters like 'lan-hosts' appear across multiple tools but are not consistently typed or validated. No evidence of parameter validation, enum constraints, or dependency documentation.
Run the Playwright MCP server with integrated LAN-direct proxy management for split-tunnel VPN bypass
Find the physical adapter and prove the bypass works by comparing direct vs bound egress IPs
Copy the skill into an agent's skills directory for discovery by Claude Code or Codex
Run the LAN-bound HTTP/CONNECT proxy whose outbound sockets are bound to the physical adapter's IP
Show which network a host would leave by (local loopback, ISP via binding, or tunnel)
No visible JSON Schema input definitions for any tool. Parameters documented only as free-form strings in descriptions. Cannot verify type constraints, required fields, or enums.
Tool names are generic and do not follow verb_noun convention. 'browser', 'proxy', 'detect', 'route' lack clarity on what action is performed. 'install-skill' is ambiguous, does it install, register, or copy?
Destructive operations (proxy, install-skill) marked WRITE but lack confirmation patterns, dry-run options, or error recovery guidance. No documentation of what happens on failure or how to undo.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 29 | 2026-07-28+ | v2 |
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. Responses are unstructured.
Tool definitions are inferred from CLI argument parsing (bin/lan-direct.mjs) rather than explicit MCP tool registration. No evidence of MCP server initialization, tool list endpoint, or schema export.