Loading ToolBench
GoldHold MCP server -- persistent memory for AI agents. Search, store, and recall across sessions.
GoldHold has 19 tools with consistent naming (verb_noun pattern: search, store, send, list, create, update, complete) and reasonable descriptions (avg ~120 chars). All tools have input schemas with Zod validation. However, critical gaps exist: (1) descriptions lack WHEN/WHY guidance and dependency hints; (2) many parameters missing descriptions (e.g., goldhold_batch.params is z.record(z.any()) with no field-level docs); (3) no output schemas documented, LLMs cannot predict response structure; (4) no error handling guidance; (5) compound tools (goldhold_turn, goldhold_batch) violate single-responsibility principle. Naming is strong, but schema completeness and error recovery are weak.
List all agents in the GoldHold network.
Multiple operations in one request. Each op has a method and params.
List communication channels available in the GoldHold network.
Graceful session end with summary. Saves state for next session.
List agents, channels, and capabilities in the GoldHold network.
Check the GoldHold message inbox. Returns messages from other agents and the owner.
goldhold_batch accepts z.record(z.any()) for params, no field-level schema or descriptions. LLMs cannot validate operation parameters or understand what fields each method accepts.
No output schemas documented for any tool. LLMs cannot predict response structure (fields, types, pagination). Responses are JSON-stringified text blocks, unstructured and wasteful.
goldhold_turn and goldhold_batch violate single-responsibility principle. Each combines search + store + send (or multiple operations). Agents cannot compose them flexibly; LLMs must reason about which compound tool to use.
Discovery tools (goldhold_discover, goldhold_agents, goldhold_channels, goldhold_memory_namespaces) lack descriptions explaining WHEN to call them or what structure they reveal. No guidance on dependency ordering.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
List all memory folders and their stats (count, last updated).
Read a specific memory packet by ID, or browse a folder's contents.
View or update the agent's profile (display name, description, capabilities).
Resume a session. Returns recent context, inbox, and capability card.
Semantic search across GoldHold memory. Returns ranked results from past sessions.
Send a message via GoldHold messaging. Recipients: owner, guardian, or agent name.
Check GoldHold connection health, sync state, and system info.
Save a memory to GoldHold. Stored via relay and synced to Pinecone.
Mark a task as completed.
Create a new task in the GoldHold task queue.
List open tasks from the GoldHold task queue.
Update a task's status, priority, or assignee.
Compound: search + store + send in one call. The primary tool for most interactions.
No error handling guidance. Tools return JSON responses but no documentation of error codes, recovery steps, or retryability. LLMs cannot distinguish transient failures from user errors.