Per-agent scoped MCP tool proxy with credential isolation and audit logging
Four tools with clear verb-noun naming (get_key, list_keys, set_key, delete_key) and consistent scoping semantics. All have descriptions and input schemas with type definitions. However, descriptions are brief (10-50 chars), parameter descriptions lack detail on constraints and formats, and output schemas are not documented. No error handling guidance, no examples of what happens on failure, and no indication of return types beyond docstring hints. The scoping mechanism is well-designed but underdocumented in tool descriptions themselves.
Delete a key (scoped to agent namespace).
Get a value by key (scoped to agent namespace).
List keys matching a pattern (scoped to agent namespace).
Set a key-value pair (scoped to agent namespace).
Tool descriptions are too brief (10-50 chars). Descriptions like 'Get a value by key (scoped to agent namespace).' lack context on when to use the tool, what it returns, and error conditions. LLMs need 50-200 char descriptions to make confident selection decisions.
Parameter descriptions lack constraint details. 'Key name. Automatically prefixed with {agent_id}:' does not specify format, length limits, or character restrictions. 'Glob pattern (e.g. "session:*")' includes an example value that LLMs may reuse literally instead of adapting to context.
Output schemas are not documented. Tool docstrings do not specify return types, structure, or fields. For get_key, it's unclear whether None is returned as null, empty string, or an error. For list_keys, no documentation of the list structure, pagination, or total count.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | <=2025-11-25 | v2 |
No error handling guidance. Descriptions do not explain what happens on failure (e.g., Redis connection lost, key not found, TTL invalid). LLMs receive no recovery hints and cannot distinguish retryable from fatal errors.
set_key TTL parameter lacks validation constraints. Description says 'Time-to-live in seconds. 0 means no expiry.' but does not specify minimum/maximum values, whether negative is allowed, or what happens if TTL exceeds Redis limits. Unbounded numeric parameters invite invalid LLM inputs.