Loading ToolBench
MCP server that exposes AIIA's persistent AI memory, knowledge base, and reasoning capabilities as tools for Claude Code. Provides access to ChromaDB-backed knowledge search, structured memory persistence, session management, and production monitoring.
AIIA has 13 tools with reasonable naming (verb_noun pattern: aiia_ask, aiia_remember, aiia_search) and descriptions present for all tools. However, critical gaps undermine quality: (1) Parameter descriptions are sparse or missing entirely, most tools lack per-parameter documentation explaining constraints, formats, or valid values. (2) Output schemas are not documented anywhere in the provided code, LLMs cannot infer what fields to expect from responses. (3) Error handling is absent, no guidance on recovery, retryability, or actionable error messages. (4) Several tools combine multiple concerns (aiia_session_end bundles summary + decisions + lessons; aiia_ask mixes search + reasoning + optional DeepSeek). (5) Tool descriptions average ~150 chars but lack WHEN/WHY context, e.g., aiia_ask vs aiia_search distinction is unclear without trying both. Baseline: 549 production tools average 194 chars for descriptions and 100% have param annotations; AIIA has ~60% param coverage and 0% output schema documentation.
Ask AIIA a question. She searches her knowledge base (5,500+ indexed docs from the AIIA repo) and her persistent memory, then reasons with her local LLM (llama3.1:8b-instruct-q8_0) to give a grounded answer.
Log a story/idea to the roadmap backlog with deduplication. Adds items to the product backlog.
Production health check via Command Center. Shows status of all monitored services.
Score and rank backlog stories using 5-filter framework. Provides prioritized backlog view.
List AIIA-prepared branches ready for human-in-the-loop coding. Shows available work branches.
Teach AIIA a fact so she remembers it permanently. This is stored on the Mac Mini and persists across all sessions forever.
No output schemas documented. LLMs cannot infer response structure, forcing them to guess what fields are returned and breaking downstream tool chaining. All 13 tools lack documented return types.
Parameter descriptions missing or incomplete. 'depth' in aiia_ask lacks format/enum guidance ('fast', 'deep', 'recursive' are mentioned but not formalized). 'filters' in aiia_prioritize_backlog has no description of expected structure. LLMs cannot validate inputs without explicit constraints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Save work-in-progress state for next session. Preserves current work context.
Fast vector search across AIIA's knowledge base — no LLM reasoning, just returns the most relevant document chunks. Much faster than aiia_ask.
Record session summary + decisions + lessons at the end of a work session. Persists to AIIA's memory for future context.
Load context from past sessions, WIP, decisions at session start. Provides continuity across work sessions.
Check AIIA's health — knowledge doc count, memory stats, model info, and whether she's online. Quick way to verify the Mac Mini connection.
Query today's token usage and costs. Provides visibility into LLM consumption.
Quick 'catch me up' on recent work, decisions, open tasks. Provides a brief summary of recent activity.
No error handling guidance. Tools provide no recovery hints, retryability classification, or actionable error messages. If aiia_ask fails to find docs or reasoning fails, LLM has no guidance on next steps.
Tool composition issues: aiia_session_end combines three concerns (summary, decisions, lessons) in one tool. aiia_ask mixes search + reasoning + optional external model routing. Split into single-responsibility tools for better composability.
Ambiguous tool selection: aiia_ask vs aiia_search distinction unclear from descriptions alone. Both search knowledge base; aiia_ask adds reasoning but description doesn't explain when to prefer one over the other. LLMs will guess and waste calls.