CloudBase AI Toolkit
Let AI coding tools build, operate, and deploy a CloudBase backend from chat.
- Source repo
- TencentCloudBase/CloudBase-AI-Toolkit
- Stars
- ★ 1.1k
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 59/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specificCodex · Claude Code
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A full-stack developer using Codex or Claude Code wants an agent to add login, a database, permissions, and deployment to a Todo application.
- Not a fit if
- Teams that need a backend outside Tencent CloudBase
- Teams that require fully offline, local-only data
- Teams unwilling to manage Tencent Cloud credentials
- Source review
- 59/100 · Major gaps
What does this agent do, and when should you use it?
CloudBase AI Toolkit is Tencent Cloud's integration layer between AI coding environments and the CloudBase backend platform. The repository combines a Plugin, Agent Skills, Hooks, and an MCP Server: the Plugin installs the stack, Skills guide implementation patterns, and MCP performs operations against real cloud resources. It publishes the @cloudbase/cloudbase-mcp npm package and supports local npx execution, Tencent-hosted HTTP access, and a self-hosted Cloud Mode. Its scope includes PostgreSQL and document databases, authorization rules, authentication, object storage, cloud functions, Cloud Run, static hosting, logs, and deployment. Code and deployed resources target the user's own CloudBase environment, while sensitive actions still require human confirmation. It is a strong fit for developers who want Codex, Claude Code, Cursor, or another supported coding tool to deliver a working backend, but it remains tied to the CloudBase service ecosystem.
After connection, CloudBase Skills guide the application structure and @cloudbase/cloudbase-mcp logs in to the user's CloudBase environment to operate resources. The MCP layer can query and change PostgreSQL or document-database data, establish data models and CRUD flows, and configure permissions and security rules. It can also author, deploy, invoke, and debug cloud functions or Cloud Run services; configure login methods and object storage; publish static sites or Mini Programs; inspect logs; and redeploy. The Plugin bundles the MCP Server, Agent Skills, and Hooks, while MCP can be configured independently when only operational access is needed. Local mode runs through npx and includes local-file upload and template features. Hosted HTTP removes the local Node.js requirement but omits some local-file capabilities, while self-hosted remote deployments can set CLOUDBASE_MCP_CLOUD_MODE=true or MCP_CLOUD_MODE=true to disable local filesystem and process tools.
- A full-stack developer using Codex or Claude Code wants an agent to add login, a database, permissions, and deployment to a Todo application.
- A WeChat Mini Program team needs AI-assisted setup of CloudBase data, realtime behavior, authentication, and publishing.
- A developer working in Cursor, WindSurf, or another MCP client wants to inspect data, manage functions, read logs, and redeploy without leaving the conversation.
- A team supporting several AI IDEs wants one Plugin installation to provide the MCP Server, Skills, and Hooks instead of wiring each component separately.
- An organization needing centralized access can use Tencent's hosted HTTP MCP, while a team requiring infrastructure control can run the server in Cloud Mode.
- A developer who wants CloudBase implementation guidance without granting resource access can install and use the Skills separately.
How do you install or deploy this agent?
You need a CloudBase environment and an AI tool that supports Plugins, Skills, or MCP. Local MCP execution requires Node.js 18.15.0 or newer. For an Open Plugin Spec tool, run npx plugins add TencentCloudBase/cloudbase-plugin. For the CloudBase AI CLI, run npm i -g @cloudbase/cli && tcb ai. For manual MCP setup, configure a server whose command is npx and whose arguments are ["@cloudbase/cloudbase-mcp@latest"]. For hosted HTTP access, use https://tcb-api.cloud.tencent.com/mcp/v1?env_id=<env_id> and provide X-TencentCloud-SecretId and X-TencentCloud-SecretKey headers. A remotely self-hosted server should set CLOUDBASE_MCP_CLOUD_MODE=true or MCP_CLOUD_MODE=true to turn off local filesystem and process tools.
How do you use this agent?
Once installed and connected, enter Login to CloudBase in the AI tool and select an existing, healthy CloudBase environment. A documented first end-to-end prompt is: Use CloudBase Skills to build a todo app with login, database and permissions, then deploy. Skills guide the code and architecture, while MCP creates or changes database, authentication, permission, and deployment resources. Verify the resulting data and APIs in your own environment rather than treating generated local files as completion. If login reports that the environment does not exist, confirm its existence and health in the CloudBase console, then sign in again and choose the correct environment. Review and confirm critical operations such as data changes, permission updates, and deployments before execution.
What are this agent's strengths and limitations?
- The Plugin installs the MCP Server, Agent Skills, and Hooks together, reducing repeated setup across compatible AI coding environments.
- It covers an end-to-end CloudBase backend workflow—database, auth, storage, compute, logs, and deployment—instead of stopping at local code generation.
- Local
npx, Tencent-hosted HTTP, and self-hosted Cloud Mode provide distinct choices for local capabilities, managed operation, or infrastructure control. - The documented integration list spans Codex, Claude Code, Cursor, GitHub Copilot, Gemini CLI, and numerous other coding tools.
- Deployments go to the user's own CloudBase environment, and local-mode code does not need to leave the machine until deployment.
- Core operations depend on Tencent CloudBase; moving to another backend provider would require replacing the database, auth, function, and deployment integrations.
- The complete local mode requires Node.js 18.15.0+, while hosted HTTP avoids that dependency at the cost of some local-file features.
- Hosted access requires a CloudBase environment ID and Tencent Cloud Secret ID/Secret Key, creating a credential-management responsibility for teams.
- Sensitive actions still need human confirmation; the Toolkit supplies capabilities and workflows, not security judgment.
- The reported CloudBase-versus-VM evaluation is limited to a specific model, agent framework, task, and controlled setup, so it is not general evidence for every workload.
How does this agent compare with similar options?
Unlike Vercel or Netlify, which the source characterizes as focusing on frontend or container delivery, this Toolkit centers on AI-operated backend building blocks such as databases, authentication, functions, and storage. The repository also describes a controlled comparison with a traditional self-managed cloud VM for the same Todo task: CloudBase performed more favorably on completion latency, token usage, and tool-call count under that specific model and task setup, but the result is explicitly not presented as universally generalizable.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| CloudBase AI Toolkit This agent | 59 · Major gaps | Agent plugin / skillFreemium | ★ 1.1k | today | TypeScript | Codex · Claude Code |
| pg-aiguide: PostgreSQL Expertise for AI Coding | 33 · Major gaps | CLIFree | ★ 1.9k | 11d ago | Python | ChatGPT · Codex · Claude Code |
| NitroStack | 23 · Major gaps | CLIFree | ★ 2.5k | 1d ago | TypeScript | ChatGPT · Claude Code |
| Fulling | 48 · Major gaps | Web appFree + model costs | ★ 2.4k | 2mo ago | TypeScript | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The workflows use explicit @claude triggers, read-only repository access, and a default-off boolean gate for cross-repository commits, providing ordinary confirmation and external-effect controls. However, the purpose of OIDC write access, CloudBase/MCP data flows, secret-handling boundaries, and product-level least privilege are not fully documented. Most dependencies and Actions are pinned or security-overridden, but the global @cloudbase/cli@latest installation and an npm install without a lockfile weaken supply-chain reproducibility. Pre-commit diff checks exist, while product-level undo or recovery procedures do not. The MIT notice and Tencent CloudBase attribution are clear, although publisher identity remains unverified by the stated registry.
Concrete tests cover build consistency, compatibility projections, skill mirroring, version sourcing, marketplace metadata, and rejection of invalid targets, supporting full static self-consistency credit. Dependency availability is only partial because tests require local NVM, zsh, and Node 24 and silently skip when unavailable, while workflows separately use Node 18 and 22. Missing-version and invalid-target cases have clear errors, but broader runtime and network failure messages are not shown.
The evidence covers web, mini-program, authentication, UI, specification workflows, and several editor projections, giving broad scenario coverage; builders also reject unsupported publication targets. Generated routing tables and concrete activation vocabulary are test-enforced, earning strong trigger-precision credit. Environment support includes Claude, Cursor, CodeBuddy, Kiro, and MCP, but divergent Node versions, shell/NVM assumptions, and size constraints prevent full marks.
Source/generated separation, compatibility projections, marketplace manifests, and a single metadata source show sound information architecture; naming mappings and semantic version formatting are test-protected. The supplied README-EN.md only redirects to an absent README.md, so installation notes cannot be credited; examples and FAQ support are only indirectly suggested by build tests and the examples-repository workflow. Limitations appear only in comments, skip conditions, and a TODO. The MIT license is complete. Package and skill versioning are present, but no changelog is evidenced. A copyright owner is named, while maintainers, support channels, and update responsibility are not.
The toolkit produces skill packages, editor compatibility files, MCP configuration, marketplace manifests, and example archives whose structure and metadata are tested, making the outputs practically usable. Projecting one skill source across multiple agents and editors supplies clear marginal value. Scores are reduced because no user outcomes, coverage measurements, or adoption evidence are supplied, and deployment cost, API expense, latency, and maintenance burden are undocumented.
Tests trace many claims to exact files, fields, generated paths, versions, and activation terms, providing strong claim traceability. The package manifest, workflows, and several test suites corroborate the build and publication structure, but the main README, skill bodies, implementation scripts, and execution results are absent. Assertions are generally distinguishable from explanatory comments, although capability and marketplace-summary claims are still validated mainly as expected strings rather than independently established facts.
- Do not treat the static test code as evidence that tests ran or behavior was independently verified; no code was executed for this assessment.
- Before enabling production credentials or writes, verify the OIDC purpose, CloudBase/MCP data destinations, secret exposure, and each tool's effective permission scope.
- Pin @cloudbase/cli, add a lockfile for example-build dependencies, and unify or document the Node 18/22/24 compatibility matrix.
- Cross-repository pushing is default-off, but when enabled it uses a PAT and git add .; confirm branch protection, token least privilege, and recovery procedures.
- Provide the main README, installation guidance, known limitations, changelog, and explicit maintenance and security contact paths.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
这一版带来三张 MCP Apps 面板（实验性、默认关闭），并把删除前的确认扩展到 SQL 与云 API 的删除动作。 🎉 新功能 MCP Apps 面板（实验性） - 三张卡：起步（连接状态 + 环境列表，点一行就切过去）、环境用量（资源点、套餐、分模块、指标曲线）、部署产物（访问地址、一键打开、重新部署） - 默认关闭，CLOUDBASE_MCP_UI=on 打开；不设、off 或拼错都按关闭处理，工具返回始终是普通 JSON 文本 - 环境列表默认列前 5 个并说明「共 37 个环境」，展开在卡片内滚动；服务端只给 20 个时明说「还有 17 个没列出来」 - 复制环境 ID 与访问地址后给 1.5 秒「已复制」 破坏性操作确认 - 借 MCP 的 elicitation（表单征询）在删除前停下来问一句：删云函数、集合、文档、存储文件、CloudRun 服务、MySQL 实例，对话框点名目标 - 破坏性 SQL（DELETE / DROP / TRUNCATE / ALTER … DROP）同样先暂停 - callCloudApi 里成对的删除动作（DeleteFunction、DeleteTable、DestroyEnv 等）纳入同一套确认 …View on GitHub ↗ -
Release notes excerpt
本次发布把「第一次用」这条路铺平：initialize 自带使用说明、远程接入默认走 OAuth，另外修了两处 skill 与工具描述的问题。 🎉 新功能 上手引导 - initialize 响应现在带服务端 instructions：先调一次 auth(status)（只读、无参数）、登录与绑定环境的顺序、可直接照抄的问法，以及一句「不要凭记忆猜 API，先读 Skill 或查知识库」 - 没有环境时，提示国内站新用户可以免费开通一个体验环境 - 同样几句也写进 auth 的工具描述，不注入 instructions 的客户端一样看得到 🐛 问题修复 - 小程序 skill 默认走小程序 CLI 路径，不再静默降级到手工流程 - searchKnowledgeBase 描述里的 skillName 示例改为参数枚举内可接受的值，不再出现「照抄示例却填不进去」 📚 文档更新 - 远程 MCP 配置示例改为 OAuth 优先：只填地址即完成配置，配置文件里不放密钥；静态密钥改为标注「不支持 OAuth 的客户端 / CI」的第二种写法（三个 README 与 MCP 配置参考） - 官方 Cloud API 参考页按上游自动同步；首次接入的环境说明与失败出口补齐 🔧 维护与工程改进（可选阅读） …View on GitHub ↗ -
Release notes excerpt
本次发布把托管 MCP 改成每请求独立的无状态实例、收紧了工具面与技能索引的取名规则，并给每个 skill 包补上许可文件，同时修掉一批登录态、套餐取值与部署校验的问题。 ⚠️ 不兼容变更 - manageMysqlDatabase 不再提供 provisionMySQL：工具面不再代客开通 MySQL 实例。需要 MySQL 的环境请先在控制台开通，再回到工具里接入。 🎉 新功能 托管与多租户 - 托管 MCP 的每个请求使用独立 transport，客户端元数据按凭据隔离缓存，多副本或滚动发布期间同一次调用不再被拆到两个进程上 - 托管服务器默认启用的插件集与本地一致（此前会单独隐藏反馈工具） 技能与工具面 - 新增 MCP 参考 Skill：远程端点、site / region、插件开关、工具 schema 与 mcporter 用法集中在一处 - 每个 skill 包在 SKILL.md 旁自带 LICENSE.md，满足 SkillHub 等平台的许可校验 - 反馈入口改为直接打开仓库的新建 issue 页面（案例 / 复盘两种模板），不再自动拼接错误报告链接 🐛 问题修复 - 项目级凭据优先于账号级登录态，避免带上另一个账号的登录信息 …View on GitHub ↗
FAQ
What does the Toolkit cost?
Where do the code and deployed resources reside?
How should a remotely self-hosted MCP server be secured?
CLOUDBASE_MCP_CLOUD_MODE=true or MCP_CLOUD_MODE=true to disable local filesystem and process tools for remote callers. Tencent's hosted HTTP service includes this protection.