Loading ToolBench
A persistent neuromorphic brain with LLM bridge exposing brain state, concepts, habits, anomalies, and memory as MCP tools for Claude and other frontier models
14 tools with mostly complete schemas and descriptions, but significant gaps in naming clarity, parameter documentation, and error handling. Tool names lack action verbs (brain_state, brain_history, brain_concept are descriptive but not verb-led). Most parameters have types and descriptions, but many lack constraints (enums, ranges, defaults). Output schemas are not formally documented. Error handling is minimal, no recovery guidance or actionable error messages. Descriptions average ~150 chars, within baseline range, but lack WHEN/WHY context for tool selection.
Anomalies detected in current sensor readings and behavior patterns by the anomaly detector
Detailed information about a specific concept cluster: label, times seen, last seen timestamp, whether it is currently active, associated sensor profiles, and historical usage over the last 7 days
Recent experience log entries: what the brain and external systems (LLM, user, sensors) have done. Optionally filter by event kind (teach_felt, label, reward, tool_call, etc.) or time range
Query the felt-state model: current emotional state recognition with confidence, list of known felt states with teaching counts, or detailed profile of a specific felt state including its signature and when it was taught
Current habits and behavioral patterns detected by the habit miner, including context for the current hour and a list of active habits
Historical brain activity bucketed over a time range. Returns hourly or custom-interval summaries including dominant patterns, active apps, average modulators, and sleep percentage. Queries the episode log for the last N hours (default 24) with configurable step size
Tool names lack action verbs. All 14 tools use descriptive nouns (brain_state, brain_history, brain_concept) instead of verb_noun pattern (get_brain_state, query_brain_history, describe_concept). LLMs infer intent from verb-led names; these require reading full descriptions.
No output schemas documented. Tools return complex nested objects (modulators dict, trend_180s, pattern object, senses dict) but no formal schema definition visible. LLMs cannot plan downstream calls or extract fields reliably without documented return types.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Retrieve episodes from the experience log matching a query: time range, concept cluster, or app name. Returns matching episodes with timestamps, sensor context, and modulators
Full snapshot of current brain state including felt-state recognition, modulators (DA, NE, ACh, 5HT), trend signature over 180s, prediction error, current pattern/concept cluster, sensor readings (app, keystroke rate, mouse rate, idle time, microphone RMS), data source status, working memory activity, sleep mode, and age in ticks/days
Explanation of current brain state: prediction error breakdown, narrative description of full state, and optionally detailed explanation of a specific modulator (DA, NE, ACh, 5HT) including its mechanism and current drivers
Get the full current brain state snapshot including modulators, active concepts, and sensor summary
Search historical brain episodes from the last N hours. Returns a summary of what happened over a time range: top concepts, average modulators, most-used apps, and total ticks
Assign a human-readable label to a concept neuron. Use when the user tells you what a concept represents
Get the top-N most active concept neurons. Optionally filter by minimum activation
Find concepts that are strongly connected to a given concept via shared synaptic weights. Uses the expansion→concept synapse weights: concepts with similar weight profiles respond to similar sensory patterns
Minimal error handling. No recovery guidance, actionable error messages, or error classification. brain_history returns {"error": "no episode log"} with no suggestion for next steps. brain_concept returns {"error": "unknown..."} (truncated). Agents cannot self-correct.
Parameter constraints missing. hours, step_minutes, limit, min_activation lack min/max bounds. LLMs can pass absurd values (hours=-100, step_minutes=0.001, limit=1000000) that break logic or cause timeouts. No enum constraints on modulator names (da, ne, ach, 5ht, sht).
Descriptions lack WHEN/WHY context. brain_state says what it returns but not when to call it vs brain_history or current_state. brain_recall, brain_experience, and episode_search all query historical data but distinctions are unclear. LLMs waste reasoning cycles deciding between similar tools.