Loading ToolBench
MCP server for personal single-user AI agent audit/verification workbench — provides governance tools for task management, audit trails, knowledge store, cost tracking, and policy enforcement
Server has 22 well-structured tools with complete JSON schemas and consistent naming conventions (verb_noun pattern). All tools have descriptions and input schemas are properly defined using Zod. However, parameter descriptions are largely absent, the schema shows types and constraints but the source code does not include parameter-level docstrings visible to LLMs. Output schemas are not documented. Error handling is minimal (no recovery guidance or error classification). Tool composition is sound (single responsibility, clear chains), but descriptions lack the LLM-optimization guidance (WHEN to use, prerequisites, dependencies) required for A-grade quality.
Append a step (prompt/response/cost/latency) to a task. WRITES to DB.
Append an immutable audit event to the hash chain. WRITES to DB.
Approve a knowledge candidate for promotion. WRITES to DB. Atomic with audit log.
Advisory budget status check against the most specific active cost_limit policy. Returns under_budget | warning | critical | over_budget | no_policy | no_cost_data | invalid. Does NOT enforce; caller decides. Abnormal statuses are auto-appended to the audit log. READ-only from the caller perspective.
Create a knowledge candidate for HITL review (data augmentation workflow). WRITES to DB.
Create a knowledge-store item in a specific layer. WRITES to DB.
Parameter descriptions missing. Zod schemas define types and constraints, but LLMs cannot see parameter-level docstrings. E.g., 'project' in list_tasks lacks explanation of what it filters; 'complexity' in preflight_check lacks guidance on when to use each enum value.
Output schemas not documented. Tools return data but LLMs have no schema reference for response fields. E.g., list_tasks returns a list but the response structure (fields, types, pagination) is not visible in tool definitions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Create a new governance task. WRITES to DB.
Aggregate token cost and latency over a time window. Supported periods: 1h, 24h, 7d, 30d, mtd (month-to-date). READ-only.
List knowledge items in draft or candidate status pending HITL review. READ-only.
Get a single playbook by numeric id or slug. v0.1 supports single lookup only; list/search coming in v0.2.
Get a single task by id. READ-only.
List governance tasks with optional filters (project, status). READ-only.
Nominate a knowledge item for promotion to the next layer. WRITES to DB.
DIAGNOSTIC read-only check of what a proposed action would trigger (policy + HITL). Its result is advisory — execution tools re-validate server-side in atomic transactions. Safe to call with no side effects.
Promote a candidate knowledge item to the incident layer. WRITES to DB. Atomic with audit log.
Reject a knowledge candidate. WRITES to DB. Atomic with audit log.
Report an incident caused by AI agent behavior. WRITES to DB.
Revoke a previously approved knowledge candidate. WRITES to DB. Atomic with audit log.
Full-text search over the knowledge store. READ-only.
Return the last N audit events (default 20). READ-only.
Change the status of an existing task. WRITES to DB.
Re-compute the audit hash chain and report any broken links. READ-only.
Error handling lacks recovery guidance. No tool descriptions explain what to do if a call fails (e.g., 'If task not found, try list_tasks() to find the correct ID'). Errors are not classified as retryable, user-fixable, or fatal.
Tool descriptions lack LLM-optimization guidance. Descriptions state WHAT tools do but omit WHEN to use them, prerequisites, or dependencies. E.g., 'preflight_check' description does not explain that it is advisory and execution tools re-validate.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools are marked READ_ONLY or WRITE in comments but not via MCP tool annotations, limiting agent reasoning about safety and retry behavior.