Loading ToolBench
Shared Memory MCP Server for coordinated multi-agent workflows with token-efficient context management
The server presents 14 well-structured tools with complete JSON Schema definitions and descriptions. Tool naming follows verb_noun conventions consistently. However, several critical gaps reduce the score: (1) parameter descriptions are present but many lack detail about constraints, valid ranges, or format expectations; (2) output schemas are NOT documented, callers don't know what fields to expect from responses; (3) error handling descriptions are absent, no guidance on recovery paths or error classification; (4) no parameter descriptions explain dependencies between fields (e.g., 'result' param in update_work_status only applies when status='completed'); (5) security patterns not visible, no audit logging, permission checks, or validation guidance shown in the code sample.
Add a discovery that other workers can benefit from
Block until a dependency output is available
Claim a specific work unit for execution
Create a new agentic session with shared context for coordinator and workers
Declare output keys that a worker will produce
Get detailed information for a specific context section
Get incremental updates since a specific version
Output schemas not documented. Tool descriptions do not specify what fields the caller will receive in the response. This forces LLMs to guess at response structure and blocks downstream tool chaining (e.g., does get_session_info return worker_ids? Does it return the full context or just metadata?).
Parameter descriptions lack constraint details. Many parameters have descriptions like 'Session ID' or 'Worker ID' but do not specify format (UUID? alphanumeric? length bounds?). Enum parameters are documented but free-form string params like 'section' in expand_context_section lack guidance on valid values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Get discoveries made since a specific timestamp
Get information about an agentic session
Get compressed context for a specific worker
Publish an output that workers are waiting for
Publish available work units for workers to claim
Update the status of an agentic session
Update the status of a work unit
No error handling guidance visible in tool descriptions. Callers don't know when tools fail, what errors are retryable, or how to recover. E.g., claim_work_unit likely fails if the unit is already claimed, but the description does not mention this or suggest a recovery path.
Undocumented parameter dependencies. In update_work_status, the 'result' parameter is optional and only meaningful when status='completed'. This conditional requirement is not stated in the description. Similarly, in create_agentic_session, codebase_files and requirements are optional but have nested required properties that are not explained.
No idempotency or confirmation pattern for destructive/write operations. Tools like claim_work_unit, update_work_status, and publish_output modify shared state but descriptions do not clarify idempotency guarantees or confirm-before-execute behavior. An agent retrying due to ambiguous network failure could claim the same work unit twice.
No pagination or result limiting visible. Tools returning lists (e.g., get_discoveries_since likely returns an unbounded array) do not document pagination parameters (limit, offset/cursor) or maximum result counts. Returning thousands of discoveries could exhaust context.
Missing tool composition guidance. When an LLM uses claim_work_unit, what IDs must it already have? Does it need to call publish_work_units first? The descriptions do not explain prerequisite calls or the typical workflow order.