// the find
TencentCloudBase/CloudBase-AI-Toolkit
Backend for AI coding agents on CloudBase — database, auth, functions via Plugin, Skills & MCP.
An MCP server plus agent skills layer that connects Tencent Cloud's CloudBase platform (managed database, auth, functions, storage) to AI coding tools like Cursor, Claude Code, and a dozen others. The Plugin installs the full stack, Skills constrain how the AI generates code against CloudBase patterns, and MCP lets the AI operate your actual environment from the conversation. Primarily aimed at developers doing vibe-coded apps on Tencent Cloud, especially WeChat Mini Programs.
The three-layer architecture (Plugin → Skills → MCP) has a real logic to it: Plugin handles setup, Skills encode platform-specific patterns the AI should follow, MCP operates live resources. That separation means you can use Skills alone for knowledge constraints without needing MCP running. The Skills files are unusually detailed — per-domain checklists, auth/RLS pattern references, function runtime gotchas — which is more useful than typical 'just ask the AI' approaches. Multiple MCP connection modes (local npx, hosted HTTP, self-hosted cloud mode with local file tools disabled) with a clear rationale for when each applies shows someone thought about the security surface. Broad IDE coverage with per-tool config snippets that are maintained as the ecosystem has fragmented.
Total Tencent Cloud lock-in — the Skills encode CloudBase-specific security rules, CloudBase-specific function runtimes, CloudBase-specific database permissions. Code generated under these Skills isn't portable; migrating away means rewriting the backend layer, not just swapping a config. The repository structure is a mess: the same skill files are duplicated verbatim across `.agents/skills/`, `.claude/skills/`, `.codebuddy/skills/`, `.workbuddy/skills/` — four copies of the same content, which will drift. The benchmark claim in the README ('more favorable in completion latency, token usage, and tool-call count') comes with no numbers, no baseline, and a disclaimer that it 'should not be generalized beyond that scope' — it's a non-claim dressed as evidence. WeChat Mini Program content is woven throughout the skills and docs, so non-WeChat developers will hit irrelevant noise constantly with no clean way to filter it out.