Gemini (Antigravity)용 GLM-4-plus 협업 MCP 서버 - 25개 도구를 통한 파일/Git/코드/네트워크/메모리 관리
Server has 25 tools with visible schemas and descriptions, but quality is inconsistent. Descriptions are present but often minimal (10-50 chars), falling below the 50-200 char LLM-optimized baseline. Many tools lack parameter descriptions entirely (e.g., glm_schedule_task has only 'action' with minimal guidance). Schemas use Pydantic models (good), but output schemas are not documented. Error handling is absent, no recovery guidance, no categorization of retryable vs fatal errors. Tool naming follows verb_noun convention (positive), but composition issues exist: glm_file_rollback, glm_memory_delete, and glm_shell_exec are destructive but lack confirmation patterns. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite MCP 2026-07-28 support. Security concerns: glm_shell_exec and glm_code_run accept arbitrary input with only whitelist-based mitigation, no input validation visible in source. glm_db_query accepts raw SQL with 'read_only' flag but no injection guards evident.
작업 로그 조회.
GLM에 원시 프롬프트를 직접 보냅니다. (Raw Mode)
GLM-4-plus에 질문을 위임합니다. (지능)
Python 코드 실행 (샌드박스).
DB 쿼리 (SQLite).
디렉토리 목록 조회.
새 파일을 생성합니다.
파일을 삭제합니다 (백업 보관).
Destructive tools lack confirmation/dry-run patterns. glm_file_delete, glm_file_rollback, glm_memory_delete, glm_shell_exec, and glm_git_commit are irreversible but have no confirmation step or dry-run mode. Agents can accidentally destroy data.
No error handling or recovery guidance. Tools return success/failure but do not categorize errors as retryable, user-fixable, or fatal. No actionable error messages (e.g., 'Invalid query syntax. Expected: SELECT ... WHERE ...'). LLM cannot self-correct.
Tool descriptions are too brief (10-50 chars). Baseline is 50-200 chars for LLM optimization. glm_memory_list ('메모리 목록.'), glm_schedule_task ('작업 예약 관리.'), and glm_grep lack context on WHEN to use them vs similar tools. LLM cannot disambiguate.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
파일을 수정합니다 (문자열 교체).
파일 내용을 읽습니다.
파일을 복원합니다.
Git 커밋.
Git 변경 사항 비교.
Git 커밋 이력 조회.
Git 상태 조회.
파일 내용 검색 (grep).
HTTP 요청.
이미지 분석 (Vision).
메모리 삭제.
메모리 조회.
메모리 목록.
메모리 저장.
작업 예약 관리.
제한적 쉘 명령 실행 (화이트리스트 기반).
웹 검색 (DuckDuckGo).
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite MCP 2026-07-28 support. Clients cannot infer which tools are safe to retry, which modify state, or which are read-only. Breaks agent planning.
Security: glm_shell_exec and glm_code_run accept arbitrary input with only whitelist-based mitigation. No visible input validation, sanitization, or injection guards in source. glm_db_query accepts raw SQL with 'read_only' flag but no SQL injection prevention evident.