Minimal MCP stdio server for Agent Lexicon. A deterministic terminology layer for AI agents with tool-call safety checks backed by terminology resolution.
Agent Lexicon MCP provides 6 tools with complete JSON schemas and descriptions. Naming follows verb_noun convention (resolve_term, check_language, guard_tool_call, find_evidence, submit_proposal, get_snapshot). All tools have descriptions (34-89 chars, baseline 194 avg). Schemas include required fields and type definitions. However, descriptions are terse and lack WHEN/WHY context. Parameter descriptions are minimal or absent in several tools (e.g., 'scopes' array lacks explanation of scope semantics). Output schemas are not documented. Error handling is present but not recovery-focused. No tool annotations (readOnlyHint/destructiveHint). One write tool (submit_proposal) lacks confirmation pattern.
Check whether text is known, resolved, or ambiguous in the local lexicon.
Find lexicon or workspace evidence for a term id or surface.
Return local snapshot metadata from the Agent Lexicon workspace.
Check whether a requested tool call is safe for the resolved terminology.
Resolve text against the local Agent Lexicon and report ambiguity.
Save a local review decision for a workspace candidate.
Parameter descriptions are missing or vague. 'scopes' array appears in 4 tools but is never explained, what do scopes represent? How should an LLM populate them? 'lexicon_path' is optional but no guidance on default behavior.
Output schemas are not documented. Tools return structured data (ToolGuardDecision, ResolutionStatus, etc.) but LLMs cannot see the response schema. This forces LLMs to guess what fields are available and risks downstream tool call failures.
Tool descriptions lack WHEN/WHY context. 'Resolve text against the local Agent Lexicon and report ambiguity' (resolve_term) does not explain when to call this vs check_language. LLMs cannot disambiguate similar tools without explicit guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
submit_proposal is a write tool with no confirmation or dry-run pattern. Agents could accidentally accept/reject candidates without review. No idempotency guarantee documented.
No tool annotations present. Tools lack readOnlyHint (5 are read-only) and destructiveHint (submit_proposal writes state). Annotations enable LLMs to reason about safety and retry logic.