Deploy Claude Code, Codex & Antigravity (agy) CLI agents to demolish your work before users do. Real file analysis. Brutal honesty. Per-call model pinning, conversation continuation & intelligent pagination.
Brutalist MCP presents 11 tools with identical schema and description patterns. While schemas are formally defined with proper JSON Schema structure, they suffer from critical issues: (1) ALL tools share generic, domain-agnostic parameter sets (context, workingDirectory, clis, verbose, models, clients, pagination fields) designed for a 'roast' abstraction rather than domain-specific purpose, making each tool feel interchangeable; (2) descriptions are moderate (185-200 chars) and mention analysis type but lack actionable guidance on WHEN to use each tool over others, the distinctions between roast_codebase, roast_file_structure, roast_architecture are unclear from descriptions alone; (3) parameter descriptions are present but generic ('Subset of native critics to run', 'Per-CLI model override') with no format constraints, enums for valid CLI names, or guidance on mutual exclusivity (why both 'clis' array AND 'clients' array?); (4) no documented output schema, making it unclear what downstream tools or agents should expect; (5) pagination fields (offset, limit, cursor, context_id, resume, force_refresh) appear on ALL tools identically, suggesting a framework-wide pattern rather than tool-specific design, but their interaction semantics are undocumented; (6) no error handling guidance, what happens if a critic CLI fails? Can the tool recover? Should the user retry?. Tool definitions are code-generated (src/tool-definitions-generated.ts) rather than hand-crafted, which explains the uniformity but also indicates limited tailoring. Schemas are visible and valid JSON Schema, preventing a score below 30, but the generic architecture and lack of tool-specific refinement caps this server in the Fair range.
Deploy brutal AI critics to systematically destroy your architecture. System design review, scalability assessment, component interaction analysis, technology choices evaluation, and architectural improvement recommendations.
Deploy brutal AI critics to systematically destroy your codebase. Comprehensive code analysis, architecture review, pattern detection, code smells, technical debt assessment, and refactoring recommendations.
Deploy brutal AI critics to systematically destroy your dependencies. Package manifest analysis, dependency bloat detection, security vulnerability assessment, version constraint analysis, and modernization recommendations.
Deploy brutal AI critics to systematically destroy your file_structure. Directory tree analysis, import graph analysis, coupling detection, modularity assessment, and structural improvement recommendations.
Deploy brutal AI critics to systematically destroy your git_history. Commit pattern analysis, contributor assessment, code churn detection, branching strategy review, and collaboration workflow recommendations.
Deploy brutal AI critics to systematically destroy your idea. Concept validation, feasibility assessment, market fit analysis, implementation risk evaluation, and strategic recommendations.
All 11 tools share identical parameter schemas. Parameters like 'clis', 'clients', 'models', pagination fields, and 'workingDirectory' are copy-pasted across every tool. This generic framework-driven approach obscures tool-specific purpose and makes tools feel interchangeable to an LLM. The pattern suggests the server invokes a shared orchestration layer rather than domain-specific implementations.
Descriptions are generic domain names repeated ('Deploy brutal AI critics to systematically destroy your X. [Generic analysis list].') with no guidance on WHEN to use roast_codebase vs roast_architecture, or what makes each domain-specific. At 185 - 200 chars, they meet length baseline but lack LLM-optimized context for tool selection.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 44 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Deploy brutal AI critics to systematically destroy your infrastructure. Infrastructure design review, scalability assessment, reliability evaluation, cost optimization analysis, and infrastructure improvement recommendations.
Deploy brutal AI critics to systematically destroy your product. Product-market fit analysis, user experience critique, feature prioritization review, competitive positioning assessment, and product strategy recommendations.
Deploy brutal AI critics to systematically destroy your research. Methodology evaluation, data validity assessment, conclusion support analysis, bias detection, and research quality recommendations.
Deploy brutal AI critics to systematically destroy your security. Vulnerability detection, threat modeling assessment, security pattern analysis, compliance evaluation, and security hardening recommendations.
Deploy brutal AI critics to systematically destroy your test_coverage. Test suite analysis, coverage gap detection, test quality assessment, edge case identification, and testing strategy recommendations.
No documented output schema. Callers cannot predict what fields each tool returns, making downstream integration and LLM planning impossible. For a server claiming 'pagination' support, the pagination contract (total count? next_cursor format? result structure?) is unstated.
Parameter 'clis' accepts enum values [codex, claude, agy] but the description does not explain what each CLI is, when to use each, or why you would override one over another. A user unfamiliar with the critique framework cannot reason about this parameter.
Dual parameters 'clis' (native critics) and 'clients' (named CLI clients) with note 'ADDITIVE to the native critics', but the interaction is undocumented. If both are passed, are all executed in parallel, sequentially, or merged? What is the precedence if both 'clis' and 'clients' specify 'claude'?
Pagination parameters (offset, limit, cursor, context_id, resume, force_refresh) lack type constraints and expected ranges. Is 'limit' chars or items? What format is 'cursor'? What is 'context_id' and when must it be passed? These fields are undocumented.
The 'models' parameter is an object with optional keys (claude, codex, agy) but the description does not specify: valid model identifiers for each key, whether partial overrides are allowed, or what happens if a model is unavailable.
No error handling guidance. If a critic CLI fails, times out, or returns invalid output, what does the tool return? Is the error retryable? Should the agent try a different critic? How does partial failure (some critics succeed, some fail) surface to the caller?