MCP server for the Bloom Constraint Kernel, providing tools to query agent state, list receipts, and propose intents via the Bloom API
Three tools with basic schemas and descriptions, but significant gaps in parameter documentation and output schema clarity. All tools are read-only (low risk), but descriptions lack actionable context for LLM selection. Parameter descriptions are minimal (e.g., 'The agent ID to fetch state for' is generic). No output schemas documented. Tool names follow verb_noun pattern but lack specificity about what state/receipts/quote mean in domain context. Average description length ~60 chars (below 194-char baseline). No error handling guidance or recovery paths documented.
Request a quote to see if an intent is allowed.
Fetch the latest state and observation for an agent.
Fetch receipts for an agent.
Output schemas not documented. LLMs cannot plan downstream calls or extract required fields. No indication of what 'state', 'observation', or 'receipts' structure contains.
Parameter descriptions are generic and lack actionable context. 'The agent ID to fetch state for' does not explain format (UUID? string? numeric?), constraints, or how to obtain one. Baseline: 72 chars per param; observed: ~30 chars.
Tool 'bloom_can_do' name is vague. 'can_do' does not clearly convey that it requests a quote or permission check. Should be 'request_intent_quote' or 'check_intent_allowed' for clarity.
No error handling guidance. Tools lack descriptions of failure modes, recovery paths, or actionable error messages. LLMs cannot self-correct on failures.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
Parameter 'intent_json' in bloom_can_do lacks schema constraints. Accepts arbitrary object with no type validation, format hints, or required fields documented. Invites hallucinated payloads.