Loading ToolBench
Source-first Skill and MCP workflow for URL-based website cloning, capture, rebuild, and fidelity verification.
18 tools with schemas present but significant quality gaps. Most descriptions are adequate (50-150 chars) but lack LLM-optimized guidance on WHEN to use each tool vs. alternatives. Parameter descriptions exist but are often generic ('The HTTP or HTTPS URL to...'). No output schemas documented. Error handling is absent, tools return raw exceptions rather than recovery guidance. Tool naming is verb-forward (generate_, inspect_, detect_, discover_, audit_, trace_, classify_, capture_, plan_, verify_, build_, clone_, enqueue_, list_, load_, cancel_, run_) which is good, but composition is poor: 18 tools operate on overlapping concerns (URL inspection, capture, cloning, job queue) without clear sequencing guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels (READ_ONLY vs WRITE). Schemas use proper JSON types but lack constraints (enums, min/max, patterns). No pagination support despite tools returning potentially large results.
Perform a low-risk routing and approval preflight audit without browser capture.
Build a rebuild scaffold from a capture bundle with optional persistence.
Build a reproduction bundle from a capture bundle.
Cancel a clone job by ID.
Capture a complete reference bundle including static inspection, runtime trace, and optional breakpoint variants.
Classify the clone mode based on policy, license, candidates, and source signals.
Clone a reference URL with full capture, rebuild, and optional runtime trace.
No output schemas documented. LLMs cannot plan downstream tool calls or extract required fields (e.g., what does capture_reference_bundle return? What fields enable plan_reproduction_path?). Breaks tool chaining.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite explicit risk levels. WRITE tools (trace_runtime_sources, capture_reference_bundle, clone_reference_url, enqueue_clone_job, cancel_clone_job, run_clone_job) lack destructiveHint; READ_ONLY tools lack readOnlyHint. Agents cannot reason about safety.
Parameter descriptions are generic and lack LLM-optimized guidance. 'The HTTP or HTTPS URL to...' appears in 8+ tools but does not explain WHEN to use inspect_url vs discover_embed_candidates vs audit_reference_url. No dependency hints or sequencing guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Detect runtime capabilities including Node.js, Playwright, browser availability, and supported session modes.
Discover embeddable candidate URLs from a reference URL.
Enqueue a clone job in the job queue.
Generate an embed snippet (iframe or JSX) for a given URL and framework.
Inspect a reference URL to extract metadata, frame policy, platform detection, and candidate URLs.
List clone jobs in the queue, optionally filtered by state.
Load a specific clone job by ID.
Plan the reproduction path based on site profile, candidates, and capture bundle.
Run a clone job from the queue, either a specific job by ID or the next pending job.
Trace runtime sources from a URL with browser capture, HTML, and screenshot options.
Verify fidelity between reference and candidate bundles or URLs.
No error handling or recovery guidance. Tools return raw exceptions rather than actionable errors. E.g., if a URL fails inspection, what should the agent try next? No categorization of retryable vs user-fixable vs fatal errors.
No parameter constraints (enums, min/max, patterns). E.g., 'framework' accepts 'html' or 'nextjs' but is not declared as an enum. 'timeout_seconds' and 'wait_seconds' lack min/max bounds. LLMs can pass invalid values.