Loading ToolBench
AI PR review assistant. Checks what your AI agent actually changed vs what you asked for. Deterministic, evidence-backed, no AI opinions.
Overreach has two tools with clear, well-written descriptions and proper Zod schemas. check_overreach is well-named (verb_noun pattern) with a detailed description (180+ chars) explaining what it audits and what it returns. Parameters are typed (string, enum) and described. However, the output schema is not documented, the tool returns JSON but the structure of findings, scope_creep_score, and other fields is not specified in the tool definition. The health tool is minimal but functional. No error handling guidance, no pagination, no per-tool risk annotations. Overall solid for a specialized audit tool, but lacks output documentation and error recovery patterns.
Audit a code diff against the originating natural-language prompt. Flags every out-of-scope change the AI agent made — unauthorized deps, env vars, endpoints, cron jobs, files, and features. Returns structured findings + a scope_creep_score.
Health check for the Overreach MCP server.
Output schema not documented. check_overreach returns JSON with findings and scope_creep_score, but the structure of findings array, field names, and data types are not specified in the tool definition. LLMs cannot plan downstream processing without knowing the response shape.
No error handling guidance. If diff parsing fails, language detection fails, or the prompt is malformed, the tool returns an error but does not guide the LLM on recovery (e.g., 'Try with language=auto' or 'Ensure diff is valid git output'). Agents cannot self-correct.
health tool description is generic (20 chars). 'Health check for the Overreach MCP server.' does not explain when to call it, what it returns, or why an agent would use it. Descriptions should be 50-200 chars and answer: what, when, why.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
No tool annotations. Neither check_overreach nor health declare readOnlyHint, destructiveHint, or idempotentHint. check_overreach is read-only and idempotent, these should be annotated per the current MCP spec (2026-07-28).
STDIO-only transport by default. The server supports Streamable HTTP via PORT env var, but defaults to STDIO. STDIO is not remotely accessible and cannot be used by hosted MCP clients. This caps protocol readiness at 50.