Loading ToolBench
Turn any MCP server into a curated, composable, A2A-ready agent.
Agentweld is a tool curation framework, not an MCP server itself. The 7 example tools span two demo servers with inconsistent quality. search_notes, draft_release_note, and create_followup_ticket have reasonable descriptions (50-100 chars) and schemas. However, the 'get' and 'post' tools have minimal/missing descriptions (4-20 chars), violating the 10-1024 char baseline. No tool declares output schemas. Error handling descriptions exist but lack recovery guidance. Tool naming is verb-first (good), but 'get' and 'post' are generic and ambiguous. Parameters are typed but many lack descriptions (e.g., 'post' payload). No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Average per-tool score: 42/100.
Create a follow-up ticket draft. Raises an invalid priority error for unsupported values.
Draft a release note for a feature. Fails when feature, audience, or changes are missing.
Gets.
List customer orders by status. Returns an empty list when no orders are found and raises an invalid status error for unsupported statuses.
Search demo notes by keyword and return matches. Returns an empty list if no note matches.
Prepare an invoice reminder message. Fails when the invoice id is invalid or email is unavailable.
Generic tool names 'get' and 'post' violate verb_noun convention and are ambiguous. LLMs cannot infer intent from names alone.
Tools 'get' and 'post' have missing or trivial descriptions (<20 chars). Violates 10-1024 char baseline and prevents LLM tool selection.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields for chaining.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite explicit risk classifications (READ_ONLY, WRITE). Agents cannot reason about side effects.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
Error handling descriptions mention failure conditions but lack recovery guidance. E.g., 'Fails when feature...missing' does not tell LLM what to do next.