Gitee MCP server demonstrates reasonable schema completeness with 20 tools explicitly defined using Zod schemas and proper JSON Schema conversion. However, significant gaps exist in description quality, parameter documentation, and error handling. Tool descriptions are present but minimal (mostly 10-40 characters in Chinese), providing limited context for LLM selection. Parameter descriptions are sparse, many critical parameters lack clear explanation of expected format, constraints, or dependencies. Output schemas are not documented in the code samples. Error handling is basic with limited recovery guidance. The implementation follows a consistent registration pattern via MCPServer.registerTool() with schema validation, but does not fully leverage patterns from the 54 Agentic Tool Patterns.
向 Gitee 仓库中的 Issue 添加评论
在 Gitee 仓库中创建一个新分支
在 Gitee 仓库中创建 Issue
在 Gitee 仓库中创建或更新文件
在 Gitee 仓库中创建 Pull Request
创建 Gitee 仓库
Fork Gitee 仓库
Minimal tool descriptions (10-40 characters). LLMs cannot distinguish between similar tools or understand when to select each tool. Descriptions like '创建 Gitee 仓库' (Create Gitee repository) lack context about prerequisites, side effects, or differentiation from related tools.
Parameter descriptions missing for critical fields. For example, 'refs' in create_branch has description 'Source reference for the branch, default: master' but does not explain whether it must be a branch name, tag name, or commit SHA. Multiple parameters across tools lack format constraints or examples.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 51 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 11 | - | v1 |
获取 Gitee 仓库中的特定分支信息
获取当前认证的 Gitee 用户信息
获取 Gitee 仓库中文件或目录的内容
获取 Gitee 仓库中的特定 Issue
获取 Gitee 仓库中的特定 Pull Request
获取 Gitee 用户信息
列出 Gitee 仓库中的分支
列出 Gitee 仓库中的 Issues
列出 Gitee 仓库中的 Pull Requests
合并 Gitee 仓库中的 Pull Request
向 Gitee 仓库提交多个文件
更新 Gitee 仓库中的 Issue
更新 Gitee 仓库中的 Pull Request
No documented output schemas in visible code. Tools return JSON via console output, but the structure of results (fields, types, pagination format) is not documented. LLMs cannot plan downstream calls without knowing what fields are returned.
Error handling is generic. CallToolRequestSchema handler catches errors but returns minimal context. For example, Gitee API 404 errors are passed through without guidance on recovery (e.g., 'Repository not found. Call list_repositories to find valid repo names').
No input validation error messages. When a Zod parse error occurs, the error response is JSON.stringify(error.errors), which is machine-readable but opaque to LLMs. Should return: 'Invalid status: must be one of [open, closed, progressing]'.
Pagination inconsistency. list_branches, list_issues, list_pull_requests accept page/per_page but do not document whether they return a total_count, next_cursor, or has_more flag. Without this, agents cannot reliably iterate through all results.
No tool annotations for destructive operations. merge_pull_request, delete operations (if present), and other WRITE/IRREVERSIBLE tools do not declare destructiveHint or idempotentHint. LLMs cannot reason about safety without these annotations.
Chinese descriptions and parameter names limit discoverability for non-Chinese-speaking LLMs and agents. Tool names are English, but descriptions are exclusively Chinese (e.g., '创建 Gitee 仓库'). Bilingual or English-primary descriptions would improve usability in global agent systems.