Loading ToolBench

Reading the current corpus…

the-council MCP server · D · ToolBench

the-council

Three-tier AI agent MCP orchestration system: Chancellor (Opus), Executor (Sonnet), Aide (Haiku) with Supervisor quality review

<=2025-11-25STDIO6 tools

57/100

Definition Quality71

Protocol Readiness47

Supportability40

Spec posture

Inferred effective spec <=2025-11-25·current baseline 2026-07-28

Static source inference · medium confidence · detected: Logging

Deprecated protocol patterns detected

Summary

The Council provides 6 tools with clear naming (verb_noun pattern) and reasonable descriptions (avg ~150 chars). All tools have input schemas with typed parameters. However, output schemas are not documented in the visible code, parameter descriptions lack detail on constraints/formats, and error handling guidance is minimal. Tool composition is sound (each does one thing), but descriptions could be more LLM-optimized per the 50-200 char baseline. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels declared in metadata.

Tools (6)

consult_chancellorread onlyauthsource verified73/100

Invoke the Chancellor (Claude Opus) directly for deep strategic analysis and planning. Returns a structured plan with steps, risks, and delegation guidance.

delegate_to_aideread onlyauthsource verified73/100

Invoke the Aide (Claude Haiku) for simple, well-defined tasks: formatting, data transformation, text processing, simple utilities. Fast and cost-efficient.

execute_with_executorwriteauthsource verified70/100

Invoke the Executor (Claude Sonnet) directly for plan implementation. Provide the task and optionally a Chancellor plan as context. The Executor has access to Read, Write, Edit, Bash, Glob, and Grep tools.

get_council_stateread onlysource verified68/100

Retrieve the state of a Council session by ID, or list all active sessions if no ID is provided.

get_supervisor_verdictsread onlysource verified68/100

Retrieve Supervisor verdicts for a session. Use flagged_only to surface only outputs the Supervisor flagged as needing attention.

orchestrateread onlyauthsource verified73/100

Issues (5)

HIGHtool ↗

Output schemas not documented. Tools return complex objects (CouncilSession, ExecutorResponse, SupervisorVerdict) but LLMs cannot see field definitions, types, or which fields are required. This forces LLMs to guess at response structure and risks downstream tool call failures.

orchestrateconsult_chancellorexecute_with_executordelegate_to_aideget_council_stateget_supervisor_verdicts

HIGHconstrained-input ↗

Parameter descriptions lack constraint details. 'problem' and 'task' params have minimal guidance on length, format, or complexity expectations. No mention of max input size, required keywords, or examples of well-formed vs malformed inputs.

orchestrateconsult_chancellorexecute_with_executordelegate_to_aide

MEDIUMtool-annotations ↗

No tool annotations despite clear risk levels. Metadata declares Risk: READ_ONLY or WRITE, but tools lack readOnlyHint/destructiveHint/idempotentHint annotations in schema. LLMs cannot infer retry safety or side-effect scope from names alone.

Recommendations

Document output schemas for all tools. For orchestrate, explicitly define OrchestrateResult fields (request_id: string, complexity: enum, result: string, session: CouncilSession). For get_council_state, document CouncilSession structure (phases, timestamps, agent_results array, etc.). Use JSON Schema format in tool definitions.
Add constraint details to parameter descriptions. E.g., 'problem' → 'A natural-language problem statement (10-2000 chars). Include context, constraints, and desired outcome. Avoid vague requests like "help me".' This matches the 50-200 char baseline for LLM-optimized descriptions.
Implement tool annotations in schema. Add readOnlyHint: true to get_council_state, get_supervisor_verdicts, consult_chancellor, delegate_to_aide. Add destructiveHint: true to execute_with_executor (it can modify files/state). This enables LLMs to reason about retry safety.
Add error recovery guidance. E.g., execute_with_executor: 'May fail if the Executor times out (>30s) or exceeds token limits. If this occurs, try breaking the task into smaller steps or providing more context to consult_chancellor first.' Reference pattern:recovery-guide.
Clarify session_id semantics. For get_council_state: 'If session_id is omitted, returns a list of all active sessions (max 50, sorted by creation time). If provided, returns the specific session state or null if not found.' For get_supervisor_verdicts: 'session_id is required. Returns an empty array if no verdicts exist for this session.'

Spec posture evidence

Inferred effective spec: <=2025-11-25.

Relies on Logging (deprecated) - log to stderr or use OpenTelemetry

Score history

Overall score trend

First recorded score · v2 rubric

57/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	D	57	<=2025-11-25	v2