Loading ToolBench
Install data engineering skill packs, starter packs, MCP templates, and runnable examples into a VS Code workspace. Provides data engineering capabilities for agents and AI tools.
This MCP server has 11 tools with significant definition gaps. Most tools lack proper input schemas, parameters are documented in descriptions but not in structured JSON Schema format with type definitions. Tool descriptions are present but often generic (e.g., 'Normalize customer data from JSONL format'). No tool has documented output schemas. Parameter descriptions exist but lack constraints (enums, ranges, formats). No error handling guidance. No tool annotations (readOnlyHint, destructiveHint). The server appears to be a collection of standalone Python scripts wrapped as tools, not a cohesive MCP implementation. Tool names are reasonable (verb_noun style) but schemas are the critical failure point.
Transform bronze-layer data to silver-layer by cleaning event records (lowercasing event_type, extracting core fields)
Execute a batch job with checkpoint support for fault tolerance, aggregating events by partition with manifest tracking and staging/publish gating
Install data engineering agent toolkit files into a target project, supporting multiple tool modes (cursor, claude, copilot, gemini, kiro, windsurf, opencode) with optional symlink creation
Normalize customer data from JSONL format by extracting and transforming customer fields (id to customer_id, name normalization, country uppercase conversion)
Detect and clean up abandoned staging prefixes from incomplete serverless Spark runs based on manifest age
Reconcile source customer data against normalized output by verifying row counts, business keys, and field transformations match between source and curated datasets
Restore silver-layer data from a snapshot file
No structured JSON Schema input definitions. Parameters documented in descriptions only (e.g., 'input' and 'output' as strings), but no type, format, or constraint metadata. Violates pattern:constrained-input and pattern:tool-description.
No output schemas documented. LLMs cannot infer what fields to expect from tool responses, breaking downstream tool chaining and forcing agents to guess at response structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 34 | 2026-07-28+ | v2 |
Run evaluation comparing baseline and with-skills concern coverage scores, generating a JSON report with delta and improvement status
Compare baseline and with-skills concern coverage scores from benchmark results to detect regressions
Validate dataset against a contract schema, optionally comparing against a previous contract version and using a reference time for temporal validation
Prove that resuming from a timeout produces identical publish output as a full run by executing full run, timeout run, and resume run then comparing results
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools like 'orphan_cleanup' and 'rollback_silver' perform destructive operations but lack destructiveHint annotation. Agents cannot reason about side effects.
No error handling guidance. Tools document no recovery paths, retryability, or actionable error messages. Agents cannot self-correct on failure.
Parameter descriptions lack constraints. E.g., 'older_than_minutes' (integer) has no min/max bounds; 'tool' (string) in install_toolkit has no enum of valid tool names. LLMs will pass invalid values.