Loading ToolBench
Among Us-themed AI agent council for structured multi-perspective planning — MCP server
Server has 5 tools with complete schemas and descriptions. Naming is creative but non-standard (council_* prefix instead of verb_noun). Descriptions are detailed (150-300 chars) and explain purpose well. Schemas use Zod with proper types. However, parameter descriptions lack constraint details (ranges, formats, enums). Error handling is basic, no recovery guidance or categorization. Output is text-based rather than structured JSON, forcing LLM parsing. No tool annotations (readOnlyHint/destructiveHint). council_override is a WRITE operation but lacks confirmation/dry-run pattern.
View past council decisions, including prompts, verdicts, and vote breakdowns.
View the current council roster, their archetypes, track records, and lineage. Shows which members have been executed and replaced by the reaper.
Override a past council verdict. This feeds back into member scoring — members whose votes aligned with the override get credit, improving the council over time.
Submit a prompt for the council to deliberate on. The council will research context, generate a naive plan, then four AI council members with distinct engineering perspectives will independently critique it, debate, and vote. Returns a refined plan synthesized from their deliberation. Automatically opens the Among Us sandbox visualization in your browser.
Open the Among Us sandbox visualization in your browser to watch the council deliberate in real-time.
Tool names do not follow verb_noun convention. All tools use council_* prefix (noun_noun). LLMs infer intent from action verbs, 'council_plan' is ambiguous (plan what? for whom?). Should be 'create_plan', 'get_members', 'list_decisions', 'override_decision', 'open_sandbox'.
Output is unstructured markdown text, not JSON. LLMs must parse free-form text to extract decision_id, verdict, votes, plan. Structured output (JSON with typed fields) is cheaper, more reliable, and enables downstream tool chaining.
council_override (WRITE operation) lacks confirmation/dry-run pattern. Agents can override decisions without preview. Should support a 'dry_run' parameter or require explicit confirmation before applying feedback to member scores.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
Parameter 'limit' in council_history has no constraint documentation. Should specify range (e.g., 1-100) and default behavior. Unbounded numbers risk LLM passing absurd values.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). council_override is destructive but unmarked. council_plan and council_members are read-only but unmarked. Annotations guide agent safety reasoning.