Schema-constrained MCP server: named queries, token-derived tenancy, audit middleware
Strong foundation with clear naming, comprehensive descriptions, and well-structured schemas. All 7 tools have verb-noun names (list_, run_, fetch_, update_, refund_, grant_, send_) and detailed descriptions (100-250 chars). Input schemas are complete with typed parameters and descriptions. Risk annotations (READ_ONLY, REVERSIBLE, IRREVERSIBLE) are excellent for agent decision-making. However, output schemas are not documented in the visible code, and some parameter descriptions lack format/constraint details. Error handling guidance is minimal, tools don't explain recovery paths or valid value ranges.
Fetch a page from an allowlisted host and return its text. Two things at once, and the caller should treat it as both: the text comes from outside (untrusted input) and the request itself leaves the network (an external channel a secret could ride out on). Hosts that are not on the allowlist are refused before any request is made.
Grant a person a role on this tenant. Privilege escalation: it widens who may act later, so approving it is a different decision from approving one action.
List the named queries this server will run, with their parameters. There is no free-SQL tool. A query not in this catalogue cannot be expressed, so start here rather than guessing a name.
Refund an invoice in full. Irreversible: money leaves and no call here brings it back. Expect the caller's approval gate to stop and ask a human.
Run one named query against the caller's own tenant. Results are capped and paginated on the server; `truncated` says a cap was hit and `next_offset` continues the page. Rows from a query with returns_untrusted_text=true contain customer-written text.
Output schemas not documented. Tools return results but the response structure (fields, types, pagination) is not visible in the provided code. LLMs cannot plan downstream tool calls or extract required data without knowing what fields to expect.
Parameter constraints not fully specified. 'status' in update_ticket_status lists valid values ('open, pending or closed') in description text, not as an enum schema. 'limit' in run_query lacks min/max bounds in schema. Format constraints (e.g., ticket_id pattern 'T-\d+') are documented in descriptions but not in JSON Schema.
Error handling lacks recovery guidance. Tools do not document what errors can occur, how to classify them (retryable vs. fatal), or what the LLM should do next. E.g., refund_invoice and grant_agent_access are irreversible but have no confirmation/dry-run pattern.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | B | 70 | 2025-06-18+ | v2 |
Send an email to the customer on a ticket.
Move a ticket to another status. Reversible: the result carries the call that puts the previous status back.
Pagination details incomplete. run_query mentions 'truncated' and 'next_offset' in description but these fields are not shown in the input schema. Output structure for paginated results is not documented.
fetch_url security model unclear. Description warns that 'the request itself leaves the network (an external channel a secret could ride out on)' but does not document what secrets might leak, how the allowlist is configured, or what the LLM should validate before calling.