Exposes the running browser's tools — snapshot a page, click, type, extract, search, manage tabs — over MCP, so any MCP client can drive the browser
Single tool 'tab_open' has a basic but incomplete definition. The description is adequate (67 chars, within 10-1024 range) and clearly states the action. Input schema is present with proper JSON Schema structure (type, properties, required, additionalProperties). However, the schema lacks parameter descriptions, the 'url' property has only a minimal description ('The URL to open'). No output schema is documented, leaving the LLM unaware of what the tool returns. The tool name follows verb_noun convention (good), but lacks error handling guidance, security considerations (opening arbitrary URLs is a WRITE risk), and composition context. No pagination, no idempotency markers, no recovery guidance for failures.
Open a URL in a new browser tab. Use this to open any web page.
No output schema documented. LLM cannot infer what tab_open returns (tab ID? success message? error details?). Breaks downstream tool chaining and forces the LLM to guess.
Parameter 'url' description is minimal (11 chars: 'The URL to open'). Does not specify format constraints, valid schemes (http/https only?), length limits, or what happens if the URL is invalid or unreachable.
No error handling guidance. Tool marked as WRITE risk but no documentation of what errors can occur (invalid URL, network failure, browser crash) or how the LLM should recover.
No idempotency or confirmation pattern. Opening a URL is a side effect, repeated calls with the same URL will open multiple tabs. No dry-run or confirmation step documented.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Tool definition is inferred from test code (src/test/codex-tools-live.ts), not from explicit MCP server registration. Cannot verify if this is the actual production tool definition or just a test stub.