Loading ToolBench
AI Workbench supervision, evidence capture, validation, acceptance gates, and PR-ready audit reports.
AI Workbench MCP has 7 tools with complete input schemas and descriptions, but quality is uneven. All tools have descriptions (10-100 chars, mostly adequate), and all parameters are typed with descriptions. However, naming lacks clarity, tools use 'workbench_' prefix but lack action verbs (should be 'create_run', 'select_model', not 'workbench_open_run'). Output schemas are not documented in the source. Error handling returns generic error envelopes without recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear WRITE vs READ_ONLY semantics. Composition is reasonable, tools chain logically (open_run → record_execution → validate_run → quality_gate), but parameter interdependencies are undocumented. Baseline: 549 production tools average 18-char names and 194-char descriptions; this server averages 22-char names and 80-char descriptions, suggesting brevity over clarity.
Analyze local Workbench run ledgers and write report artifacts.
Create a Workbench run folder and initial evidence artifacts.
Run the Workbench quality gate for a run directory.
Capture Goose/model response text into Workbench evidence artifacts.
Select a Workbench model tier and write model_selection.json.
Recommend an advisory Workbench policy pack from task metadata.
Run deterministic Workbench validation over a run directory.
Tool names lack action verbs. All 7 tools use 'workbench_' prefix but no clear verb (open, select, record, validate, analyze). Should be 'create_run', 'select_model_tier', 'capture_execution', 'run_validation', 'evaluate_quality', 'analyze_run_ledger'. LLMs infer intent from verb, current names force description parsing.
Output schemas not documented. No tool returns a documented schema, LLMs cannot plan downstream calls or extract fields. Error responses return generic error_envelope() with 'operation', 'code', 'message' fields, but success responses are opaque (JsonObject type alias). Baseline: 100% of A+ tools document return types.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | 2026-07-28+ | v2 |
Error handling lacks recovery guidance. _tool_error() returns {'operation', 'code': 'mcp_tool_failed', 'message': str(exc)}. No actionable next steps, no error classification (retryable vs user-fixable vs fatal), no suggestions for recovery. Pattern: error responses must tell LLM what to do next.
No tool annotations despite clear semantics. workbench_open_run, workbench_record_execution, workbench_select_model are WRITE operations; others are READ_ONLY. MCP spec supports readOnlyHint, destructiveHint, idempotentHint annotations, none present. Agents cannot infer safety/idempotency without them.
Parameter interdependencies undocumented. workbench_open_run has auto_select_policy_pack (bool) and policy_pack (str) params, mutually exclusive? workbench_select_model has validation_strength, instruction_following, complexity_score, which combinations are valid? Undocumented dependencies cause silent misuse.