Loading ToolBench
Production-grade AI coding toolkit with 70 agents, 194 skills, 97 commands, parallel multi-agent commands, semantic intent routing, self-learning memory, and a built-in MCP server (16 tools / 6 prompts / 377 resources) that bridges to Claude Desktop, LangGraph, AutoGen, CrewAI, and OpenAI Agents SDK
15 tools with clear verb-noun naming and generally good descriptions (avg 150-200 chars). All tools have input schemas with type definitions. However, parameter descriptions are sparse or missing for several tools (e.g., route_intent's top_k lacks range guidance; recall_memory's limit lacks min/max). Output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal; most tools lack recovery guidance. No tool annotations (readOnlyHint/destructiveHint). capture_memory is WRITE but unmarked. Composition is sound (single responsibility per tool), but schema completeness and error guidance fall short of A-grade standards.
Captures a new memory entry in the local memory store with problem, approach, tags, language, and project context
Returns statistics about the ECC catalog including counts of agents, skills, commands, rules, and bundles
Retrieves the full content of a specific ECC agent by name
Retrieves the full content of a specific ECC bundle by name
Retrieves the full content of a specific ECC command by name
Retrieves the full content of a specific ECC rule by name
Retrieves the full content of a specific ECC skill by name
Output schemas not documented. LLMs cannot predict response structure (e.g., what fields list_agents returns, what recall_memory result shape is). Forces agents to guess or fail on downstream chaining.
Parameter constraints missing or vague. route_intent.top_k says 'default 3, max 20' in description but no min/max in schema. recall_memory.limit lacks bounds. LLMs may pass invalid values.
No tool annotations. capture_memory is WRITE (destructive) but lacks destructiveHint. All READ_ONLY tools lack readOnlyHint. LLMs cannot distinguish safe from risky operations.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2026-07-28+ | v2 |
Lists all available ECC agents with their names, descriptions, and relative paths
Lists all available ECC bundles with their names, relative paths, and byte sizes
Lists all available ECC commands with their names, descriptions, argument hints, and relative paths
Lists all available ECC rules with their names, relative paths, and byte sizes
Lists all available ECC skills with their names, descriptions, origins, and relative paths
Returns statistics about the local memory store including total memories, tags, languages, and projects
Searches the local memory store for relevant memories based on a query string. Optionally scoped to a specific project.
Suggests the best ECC agent for a user message. Uses simple token overlap against agent name + description. This is intentionally thin — the heavy routing rule lives in rules/common/agent-intent-routing.md (also exposed as an MCP prompt). Clients should layer their own LLM judgment on top.
Error handling minimal. Tools return ok()/err() with text only. No recovery guidance (e.g., 'route_intent: message had no usable tokens' does not suggest what to do next). No error classification (retryable vs fatal).
List tools lack pagination. No limit/offset or cursor parameters. If catalog grows large, list_agents/list_skills/etc. will return unbounded results, bloating context and risking timeouts.