MCP server for Goldfish — your AI finally has a memory. Gives MCP-compatible AI tools (Claude Desktop, Claude Code, Cursor, Windsurf, etc.) access to the user's screen memory from Goldfish.
Three tools with basic descriptions but significant schema and parameter documentation gaps. goldfish_recall has a reasonable description (58 chars) and two typed parameters with descriptions. goldfish_context has minimal description (56 chars) and no parameters. goldfish_summary has adequate description (54 chars) but only one parameter. None of the tools document output schemas, which is critical for LLM planning. Parameter descriptions are present but terse. No error handling guidance, no enums for constrained inputs, no pagination support despite recall likely returning multiple results. Naming follows verb_noun pattern (good), but descriptions lack WHEN/WHY context that LLMs need for tool selection.
Smart context: recent activity, apps, current focus
Search memories by keyword/topic
Daily summary of what the user did
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., memory IDs, timestamps, tags). goldfish_recall must document what fields each memory object contains.
goldfish_context has empty input schema ({}) and minimal description (56 chars). LLMs cannot determine what this tool returns or when to call it instead of goldfish_recall. Description lacks WHEN/WHY context.
goldfish_recall accepts 'limit' parameter but no pagination support documented. If results exceed limit, LLM has no way to fetch next page. No total_count or next_cursor returned.
No error handling guidance. If query returns no results, if database is unavailable, or if date format is invalid (goldfish_summary), LLM receives no recovery hints. Descriptions do not state prerequisites or failure modes.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 48 | 2026-07-28+ | v2 |
Parameter descriptions are terse (avg ~30 chars). 'Search query to find relevant memories' lacks format guidance. 'Date for summary (YYYY-MM-DD format, default: today)' is better but does not explain what 'today' means in edge cases (timezone, DST).