SSE-based MCP server providing memory search, dream introspection, health tracking, radio/music playback, toy control, image generation, and community social features for the Rifugio personal companion system
Server has 25 tools with schemas and descriptions present, but quality is inconsistent. Most tool descriptions are adequate (50-150 chars) and follow action-verb naming (search_, get_, write_, etc.). However, parameter descriptions are often generic or missing context. Many tools lack output schema documentation. Error handling guidance is absent. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Several tools (toy_*, radio_*) have minimal descriptions. The server mixes concerns (e.g., read/write as generic database accessors) and lacks idempotency guidance for write operations. Baseline: 549 production tools average 194 chars for descriptions and 72 chars per param; this server averages ~100 chars for descriptions and ~40 chars for params.
List AI-generated sticker packs filtered by emotion/mood
Manage frequently-used AI sticker packs in resident/pinned area
Retrieve top-scored memory buckets with decay-weighted importance and arousal scoring for relationship sedimentation
Dream introspection tool - retrieve full memory content with dehydrated summaries for Claude context injection
Mark memory buckets as seen/reviewed to update activation tracking
Generate images using AI with preset configurations
Generic database accessors (read/write) lack specificity. Tool names do not convey what data is being accessed. Parameter 'table' is a free-form string with no enum constraint, inviting hallucinated table names. No output schema documented.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk classifications. Agents cannot distinguish safe reads from destructive writes without parsing descriptions. toy_stop, radio_stop, and write operations should be marked destructiveHint=true.
Toy device tools (toy_status, toy_set, toy_sequence, toy_flow, toy_wild, toy_stop) have minimal descriptions (40-50 chars). Descriptions do not explain what the device is, when to use each variant, or what happens on failure. toy_wild lacks any explanation of 'uncontrolled mode'.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | <=2025-11-25 | v2 |
Retrieve health data including cycle, sleep, steps, and heart rate metrics
Write a 'feel' type bucket to record emotional state or relationship sedimentation
List available image generation preset configurations
Post or comment on community social feed
Play music, radio stations, or audio content with bitrate control
Search for songs, radio stations, or audio stories
Stop current radio/music playback
Read from community database tables (memories, facts, conversations, etc.)
Search memory buckets using hybrid keyword and semantic search with decay scoring
Set a sleep timer for gradual audio fade-out
Generate speech output from the Companion character
Execute continuous curve/flow pattern on toy device
Execute a predefined action sequence on toy device
Control toy device with three-channel intensity values
Query current status of connected toy/relay device
Emergency stop for toy device
Activate wild/uncontrolled mode on toy device
View community social feed/moments (Pengyou Quan style)
Write or update records in community database tables
No output schemas documented for any tool. Agents cannot plan downstream calls or extract required fields. E.g., search_memory returns results but no schema shows whether results include 'id', 'score', 'content', or 'timestamp'. Baseline: 100% of A+ tools document return types.
No error handling guidance. Tools do not explain what to do if a call fails (e.g., 'If toy device is offline, check connection and retry' or 'If table does not exist, call read with table=memories'). Agents have no recovery path.