MCP server — reversibility intelligence for autonomous AI agents
The server has 4 well-named, read-only tools with good verb-noun conventions and clear input schemas using Zod validation. However, tool descriptions lack depth for LLM planning, parameter descriptions are minimal, and output schemas are entirely undocumented. Error handling exists but provides minimal recovery guidance. No structured output documentation means downstream agents cannot reliably parse or chain results.
Check if an action on a platform can be reversed. Returns reversibility level, risk score, recovery mechanism, and safer alternatives.
Get all reversibility actions for a specific platform. Free endpoint — no payment required.
List all platforms with reversibility data. Free endpoint — no payment required.
Look up detailed reversibility information for a specific platform action. Returns recovery mechanism, time window, and safer alternatives.
Output schemas are completely undocumented. Tools return JSON via fetch but no schema is declared, forcing LLMs to parse unstructured text responses. This violates the response-shaper pattern and makes tool chaining unreliable.
Tool descriptions lack depth and context for LLM selection. 'Check if an action on a platform can be reversed' is superficial, it does not explain WHEN to call this vs lookup_action, what constitutes a 'reversibility level', or what risk scores mean. LLMs cannot distinguish between check_reversibility and lookup_action from the descriptions alone.
Parameter descriptions are minimal placeholders (e.g., 'Platform name (e.g. github, slack, aws)'). They do not explain format constraints, dependencies, or how platforms relate to actions. The 'scope' and 'count' parameters in check_reversibility lack a clear relationship explanation, when should count be provided, and is it required if scope='batch'?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 54 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 46 | - | v1 |
Error responses return raw JSON with minimal guidance. A timeout or API error returns { "error": "request_failed", "message": "..." }, which gives the LLM no recovery strategy. Should it retry? Call a different tool? Check configuration?
No pagination support. list_platforms and get_platform offer no limit or offset parameters. If platforms or actions grow large, results could be truncated or cause context overflow. Best practice is to cap results at 20-50 items and document the limit in the description.