Loading ToolBench
Exposes a local kanban board to AI agents over MCP stdio. Provides tools for task management, card proposals, similarity search, and link fetching.
Server has 6 tools with complete JSON schemas and descriptions. Naming follows verb_noun convention (propose_card, find_similar, fetch_link, list_tasks, get_task, update_task). Descriptions are present and actionable (avg ~100 chars). However, several parameters lack descriptions (e.g., 'checks' array in propose_card has no description of what constitutes a valid checklist item), and output schemas are not documented. Error handling is minimal, no recovery guidance or error categorization visible. The server uses Google ADK for tool registration, which is non-standard for MCP (MCP uses native tool definitions). Tool composition is reasonable but could be tighter: propose_card and update_task both modify state without explicit confirmation patterns.
Fetch one http(s) document and return its text. Use it to read a specification or issue the user referenced by URL.
Search existing cards and import history for work that already covers a proposal. Returns cheap stubs.
Retrieve one kanban card by reference (sequence number or UUID prefix).
List kanban cards in the user's board, filtered by status and tags.
Propose one kanban card. Call this once per card; the cards are collected server-side and returned to the user, so never repeat them in your reply.
Patch one kanban card. Only provided fields change; omitted fields are preserved.
Output schemas not documented. LLMs cannot plan downstream calls or extract required fields without knowing what each tool returns.
Several parameters lack descriptions: 'checks' in propose_card (what format? markdown? plain text?), 'tags' filtering logic in list_tasks (AND or OR semantics?), 'blocked' in update_task (what does this flag do?).
No error recovery guidance. If propose_card fails due to invalid priority or effort, the LLM has no hint what to try next. Error responses should guide the agent toward self-correction.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Destructive operations (propose_card, update_task) lack confirmation or dry-run patterns. Agents can accidentally create or modify cards without explicit user approval.
Tool registration uses Google ADK (google.golang.org/adk/v2) rather than native MCP SDK. This is non-standard for MCP servers and may cause compatibility issues with MCP clients expecting standard MCP tool definitions.