Loading ToolBench
MCP server for DeepPlan — upgrade shallow IDE plans into expert-level architecture blueprints via a Council of AI Architects
Two tools with complete Zod schemas and descriptions. Tool naming is clear (verb_noun pattern). Descriptions are detailed and context-rich (200+ chars each), explaining WHEN to use each tool. However, parameter descriptions lack actionable constraints (no enums, ranges, or format specs). Output schemas are undocumented, callers cannot predict response structure. Error handling is present but generic ('Unknown error occurred'). No tool annotations (readOnlyHint, destructiveHint). Both tools are READ_ONLY but this is not declared in the schema.
Analyzes a draft architecture plan and recommends the most relevant expert personas to include in the council review. Returns a list of recommended persona names and focus areas. Use this when: - You want the AI to pick the best experts for a specific plan - The plan covers a niche domain that may benefit from specialized personas This tool does NOT run the full council — it only recommends which personas to use.
Sends a draft architecture plan to the DeepPlan backend for expert-level review. Returns structured architectural directives, edge cases, constraints, and numbered next steps that you (the IDE agent) can execute immediately. Use this when: - A plan feels too shallow or generic - Starting a complex feature or new project - You need architecture-level guidance before writing code The response includes a "Next Steps for Agent" section with numbered instructions you can follow directly. Internally, this tool runs a "Council of Architects" — 4 expert AI personas (Security, Performance, UX/DX, DevOps) analyze the plan in parallel, then a Lead Architect synthesizes their reports into a unified blueprint.
Output schemas undocumented. Both tools return text responses but callers cannot predict structure (is it JSON, markdown, plain text?). LLMs cannot reliably parse or chain results.
Parameter descriptions lack actionable constraints. 'tech_stack' accepts free-form strings with no format guidance. 'model_choice' references 'google/gemini-2.5-pro' as example but no enum or validation. LLMs will hallucinate invalid model IDs.
No tool annotations. Both tools are READ_ONLY (no side effects) but schema lacks readOnlyHint. Agents cannot distinguish safe tools from destructive ones without reading descriptions.
Error messages are generic and non-actionable. 'Unknown error occurred' and 'Error: [message]' give LLMs no recovery path. Missing error classification (retryable vs fatal) and no guidance on next steps.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Parameter 'available_personas' is an array of objects but no schema validation shown. If a persona object is malformed, error handling is unclear. No min/max array length constraints.