Reverie is a data store + command execution MCP with 20 tools, most with reasonable parameter descriptions and clear intent. All tools have non-empty descriptions (10-150 chars typically) and defined input schemas using Zod. However, several critical issues prevent a higher score: (1) Tool names are inconsistently prefixed, most use reverie_* convention, but lack clarity about whether they operate on entries, aliases, or confirm keys (e.g., reverie_alias_set vs reverie_set create scope confusion). (2) Descriptions are brief but sometimes vague, e.g., reverie_run 'Execute a stored command' doesn't explain what happens if the command fails, whether it's async, or what the output contains. (3) Output schemas are not explicitly documented anywhere in the code provided, the code shows input validation via Zod but no declared return type structures. (4) Error handling guidance is minimal, most tools don't explain recovery paths or why a call might fail. (5) Several tools combine responsibilities (e.g., reverie_export handles all three export types in a single tool with a 'type' enum; could be split). (6) Parameters occasionally lack granularity, e.g., reverie_find's 'query' parameter doesn't specify whether it's regex, substring, or fuzzy matching.

• Add explicit output schema documentation to every tool. Use TypeScript types or JSON Schema comments in the code to declare return structure (e.g., reverie_get returns {keys: string[], values?: Record<string, string>}, total: number}). This unblocks LLM output parsing.
• Expand reverie_run description to: 'Execute a stored shell command synchronously. Requires confirmation unless yes=true or a valid confirmation token. Returns {output: string, exit_code: number, stderr: string}. If command fails, check syntax with reverie_get(key, source=true).'
• Rename core tools to disambiguate entry vs alias vs confirm scope: reverie_entry_set, reverie_entry_get, reverie_entry_remove (instead of reverie_set/get/remove). Keep reverie_alias_* and reverie_confirm_* prefixes. This eliminates LLM confusion about which tool operates on what.
• Document reverie_find matching strategy in description: 'Search entries by substring match on keys and values (case-insensitive). Returns up to `limit` results with matched key paths and preview of values if values=true.'
• Split reverie_export into reverie_export_entries, reverie_export_aliases, reverie_export_confirm_keys. Each tool has one responsibility, making LLM tool selection deterministic.
• Split reverie_import into reverie_import_entries, reverie_import_aliases, reverie_import_confirm_keys for the same reason.
• Split reverie_reset into reverie_reset_entries, reverie_reset_aliases, reverie_reset_confirm_keys.
• Document confirmation token flow in reverie_run: 'For encrypted commands, call with decrypt=true to receive a one-time token valid for 5 minutes. Pass the token and yes=true in the next call to execute.'