MCP server driving a headless browser over CDP via Playwright: web automation + traffic analysis, with per-session browser isolation
Strong tool naming (verb-first, clear intent) and comprehensive input schemas with proper types and enums. Descriptions are present and actionable (avg ~120 chars). However, output schemas are not documented in the source, and several tools lack parameter descriptions for optional fields. Error handling guidance is minimal. Tool composition is sound (single responsibility, proper chaining via refs). Security posture is good (no secrets in params, API key gating). Baseline: 9 tools, all with names, descriptions, and input schemas visible.
Navigate to a URL and capture the request/response log (method, URL, status, headers, sizes, timing) plus an aggregated summary.
Reset the traffic log.
Wait for a CSS selector to be visible, then click it.
Click the element identified by a snapshot ref. Refs come from `snapshot`. Any mutation (click, type, navigate, etc.) invalidates them — re-snapshot after each mutation.
Return the currently accumulated request/response log (since the last clear/analyze). Optionally filter by substring or resource type.
Navigate the browser to a URL.
Output schemas not documented in source code. LLMs cannot plan downstream tool calls or extract required fields without knowing what each tool returns.
snapshot description mentions 'ref' and 'e3' examples but does not explain the ref format, stability guarantees, or when refs become invalid. Agents need explicit guidance on ref lifecycle.
Error handling descriptions are absent. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next (e.g., 'selector not found, try snapshot() to refresh refs').
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | <=2025-11-25 | v2 |
Capture an accessibility-tree snapshot of the current page as text, with a stable `ref` (e.g. `e3`) on every addressable node. Use refs with click_ref/type/fill/hover_ref/select_option/drag/wait_for_ref instead of brittle CSS selectors. Refs come from `snapshot`. Any mutation (click, type, navigate, etc.) invalidates them — re-snapshot after each mutation.
Type text via the keyboard, one key event at a time (fires real keydown/input/keyup listeners, unlike `fill`). With `ref`, types into that element; without it, types at whatever currently has focus. Refs come from `snapshot`. Any mutation (click, type, navigate, etc.) invalidates them — re-snapshot after each mutation.
Wait for a CSS selector to appear in the DOM before proceeding.
analyze_traffic and get_traffic accept optional 'limit' parameter (default 200) but do not document pagination strategy or whether results are sorted/filtered deterministically. Large result sets risk context window exhaustion.
type tool accepts optional 'ref' parameter but description does not clarify behavior when ref is omitted (types at current focus). Ambiguous parameter relationships invite misuse.