Model Context Protocol (MCP) bridge for OpenClaw Gateway
Server exposes 5 tools with basic Zod schemas and descriptions, but critical gaps undermine production readiness. Tool names lack action verbs (openclaw_browser, openclaw_system_run are vague about what they do). Descriptions are present but generic (10-50 chars for most). Parameters have type definitions but lack actionable constraints, e.g., 'action' enum is documented but no guidance on when to use each action. No output schemas documented. Error handling is absent: tools return raw JSON without recovery guidance. The openclaw_gateway_call tool is a dangerous pass-through with no validation or permission checks. No pagination, no result limits, no idempotency keys visible in most tools.
Unified local browser control (navigate, click, type, tabs, etc.)
Take browser screenshot
Get optimized browser accessibility tree (pruned for AI context)
Direct RPC access to Gateway
Run shell command on local machine
openclaw_gateway_call exposes raw RPC access with no validation, permission checks, or rate limiting. Accepts arbitrary method names and params, invites injection attacks and uncontrolled side effects.
Tool names lack action verbs. 'openclaw_browser' does not convey what action it performs (navigate, click, type, evaluate). LLMs cannot infer intent from the name alone.
No output schemas documented. Tools return raw JSON without field descriptions. LLMs cannot plan downstream calls or extract required data reliably.
No error handling or recovery guidance. Tools return raw JSON or stack traces without telling the LLM what to do next (retry, ask user, abort).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 48 | 2026-07-28+ | v2 |
openclaw_system_run accepts arbitrary shell commands with no sanitization. Vulnerable to command injection if LLM is tricked via prompt injection.