Loading ToolBench
A desktop application for multi-agent AI conversations with local model support, web tools, memory management, and routine automation
Blobbies defines 5 tools with basic descriptions and minimal input schemas. Tool names follow verb_noun convention (web_search, web_fetch, remember, update_memory, forget), which is good. However, descriptions are terse (10-100 chars), lack WHEN/WHY guidance, and parameter descriptions are minimal. Input schemas are present but sparse, only 'type' and 'description' fields, no constraints, enums, or validation rules. No output schemas documented. Error handling is absent, no recovery guidance or error categorization. The memory tools (remember, update_memory, forget) lack critical details: what is a memory ID? How are IDs generated? What happens on duplicate keys? Web tools lack rate-limit guidance, timeout specs, or failure modes. No tool declares permissions or security boundaries.
Delete a memory entry
Store a fact about the user or context in memory for later recall
Update an existing memory entry
Read one web page and return its text. Use when you have a specific URL: one the user gave you, or the best result from web_search. Snippets are not an answer — fetch the page before saying what it contains. HTTPS only.
Search the web and return results with titles, URLs and snippets. Use when you need to find information online.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., does web_search return URLs, snippets, or both? What fields does remember return?). This violates pattern:tool and forces LLMs to guess.
Memory tools (remember, update_memory, forget) lack critical operational details: How are memory IDs generated and returned? What happens if the same fact is remembered twice? Can updates fail if the ID doesn't exist? No error guidance or recovery paths.
Input schemas lack constraints, enums, or validation rules. web_search accepts any string as 'query', no length limits, no guidance on multi-word vs single-word queries. remember accepts any 'text', no max length, no format guidance. This invites hallucinated or oversized inputs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
No error handling or recovery guidance. What happens if web_fetch times out? If a URL is invalid? If memory storage is full? LLMs receive no actionable error messages and cannot self-correct or retry intelligently.
Tool descriptions are under 100 chars and lack WHEN/WHY context. 'Store a fact about the user or context in memory for later recall' does not explain: When should remember be called vs update_memory? What is the memory lifetime? Is it per-session or persistent? This forces LLMs to guess.