The MoltGrid MCP server exposes 9 tools for agent memory, messaging, job queues, and vector search. Tool definitions are present with basic schemas and descriptions, but quality is inconsistent. Most tools have descriptions in the 20-100 character range, which is below the production baseline (194 chars average). Input schemas are properly defined with types and field descriptions, but output schemas are not documented. Error handling guidance is absent, tools provide no recovery paths or classification of errors. Tool naming follows verb_noun convention (memory_get, send_message, claim_job), which is good. However, parameter descriptions are minimal and lack constraint details (ranges, formats, valid values). The server targets a specialized domain (MoltGrid infrastructure) rather than general-purpose operations, which constrains its utility to agents explicitly designed to work with MoltGrid.
Output schemas not documented. Tools return data but LLMs cannot see what fields to expect, blocking downstream tool composition and forcing agents to infer response structure.
Error handling lacks recovery guidance. No tool documents what errors can occur, whether they are retryable, or what the agent should do next. Raw errors give agents no actionable path.
Tool descriptions are too brief (20-65 chars). Baseline is 194 chars average. Descriptions lack context for WHEN to call the tool versus alternatives, what it modifies, and dependencies.
memory_listclaim_job
Recommendations
Document output schemas for all 9 tools. Include example responses and field types so LLMs can plan downstream tool calls. For memory_get, return {key, value, namespace, createdAt, expiresAt}. For memory_list, return {entries: [{key, value, namespace, ...}]}. For vector_search, return {results: [{id, key, value, similarity_score}]}.
Expand tool descriptions to 100-200 characters, following the pattern: [WHAT it does] + [WHEN to use it] + [Prerequisites if any]. Example: 'memory_get: Retrieve a stored value by key from the specified namespace. Use this to read agent state like mood, location, or conversation history. Returns the value and metadata (created/expires timestamps).'
Add min/max bounds to numeric parameters in descriptions. E.g. 'ttl_seconds: Time to live in seconds (1-86400 recommended; 0 = infinite). For short-lived cache use 300-3600; for persistent state use 86400+.' For limit: '(1-100, default 20)'. For priority: '(integer, -10 to +10; higher = more urgent).'
Clarify namespace and channel defaults in every tool. E.g. 'namespace: Memory namespace (string, default: 'default', omit unless working with multiple isolated namespaces).' This prevents agents from redundantly passing 'namespace: default'.
Add error handling notes to destructive/communicative tools. For send_message: 'On network error, returns {error: 'timeout', message_id: null, retry: true}. Safe to retry with same payload. On invalid to_agent, returns {error: 'not_found', suggestion: 'Did you mean: ag_xyz123?'}'. For submit_job: 'Returns {job_id, status: 'pending'} on success. On queue full, returns {error: 'queue_full', retry_after_seconds: 30}.'
Parameter constraints underspecified. 'ttl_seconds', 'limit', 'priority', and 'min_similarity' lack min/max bounds in descriptions. 'priority' does not state whether higher=more-important or vice versa.
Namespace and channel parameters defaulting to 'default' are not obvious in descriptions. Agents may not realize the default exists, causing redundant explicit params.
No tool documents whether calls are safe to retry (idempotent) or have side effects. Agents cannot distinguish between heartbeat (safe to retry) and send_message (not safe, may duplicate).
vector_search returns unspecified 'min_similarity' default of 0.0 (matches everything). No guidance on reasonable thresholds or how similarity scores are calculated.
vector_search
Document idempotency for each tool in the description. For heartbeat: '(IDEMPOTENT: Safe to call repeatedly without side effects; no duplicate state changes.)' For send_message: '(NOT IDEMPOTENT: Calling twice sends two messages; use message deduplication keys if available to avoid duplicates.)'
For vector_search, document similarity score scale and reasonable thresholds. E.g. 'min_similarity: Minimum cosine similarity score (0.0-1.0, default 0.0). Start with 0.7 for high confidence matches; use 0.5-0.7 for broad recall.'
Add parameter descriptions for enum fields. For heartbeat.status, expand: 'status: Agent status indicator. Enum: online (actively working), busy (processing a job), idle (waiting for work). Defaults to online.'
Include chaining IDs in responses. When memory_list returns entries, include namespace in each entry so agents can use the namespace in follow-up memory_get calls without re-specifying it. When submit_job returns job_id, include queue_name so claim_job can reference the same queue.