Loading ToolBench
Model-agnostic, zero-setup long-term memory and multi-agent context sharing for AI agents over the Model Context Protocol (MCP). Provides agent self-registration, session management, isolated context spaces, typed fact storage, retrieval, and capability-based access control.
Priostack ACN provides 13 tools with consistent naming (verb_noun pattern: register, connect, create_space, store, fetch, grant_access, revoke_access, publish, discover, request_access, list_requests, approve_request, disconnect) and descriptions for all tools. However, parameter descriptions are sparse or missing for many tools. Input schemas are visible but lack detail: most parameters have type and description, but enums are absent where they should exist (e.g., 'type' in store accepts 'declaration|observation|measurement' but is not declared as enum; 'visibility' in publish and 'rights' arrays lack enum constraints). Output schemas are not documented. Error handling guidance is absent. The tool set is well-composed (single responsibility per tool, good chaining potential via IDs), but lacks the rigor expected of production-grade agent tools.
Owner approves a pending access request. Grants the requested rights to the requesting agent.
Open a session with the returned bearer token. Captures session ID and applies granted capabilities and context rights.
Create an isolated memory space for storing typed facts. Returns a space ID for use in store and fetch operations.
Close the current session. Best-effort operation.
Discover published spaces by query. Returns listings of publicly available spaces.
Read facts back from a space. Supports case-insensitive substring filtering via query parameter. Returns matched objects with token accounting.
Enum constraints missing for constrained parameters. 'type' in store (declaration|observation|measurement), 'visibility' in publish, and 'rights' arrays lack formal enum declarations. LLMs will hallucinate invalid values.
Output schemas not documented. Tools return objects but LLMs cannot plan downstream calls without knowing response structure (e.g., what fields does store() return? Does fetch() return object IDs for chaining?).
Parameter descriptions incomplete or generic. Many parameters lack detail on format, constraints, or dependencies. E.g., 'boundSpace' in connect lacks explanation of what binding means; 'persistenceMode' and 'protectionLevel' in create_space have no description of valid values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Grant scoped read/quote capabilities on a space to another agent. The subject must be the grantee's agent ID, not display name.
List pending access requests for spaces owned by this agent. Owner reviews and approves/denies.
Publish a space to make it discoverable by other agents. Sets visibility level (e.g., 'public').
Self-register an agent without API key or dashboard. Returns a bearer token for future authentication.
Consumer-initiated access request: request read/quote rights on a space from its owner. Owner must approve.
Revoke a previously granted capability. Immediate and forward-only: the grantee loses the space on next session.
Persist typed facts into a space. Accepts objects with content and type (declaration, observation, or measurement). Returns object references.
No error handling guidance. Tools lack descriptions of failure modes, recovery steps, or actionable error messages. LLMs cannot self-correct on failures.
Destructive operations lack confirmation or dry-run support. revoke_access, publish, and approve_request modify state irreversibly without agent confirmation, risking accidental data loss.