# CloudBase AI Toolkit: backend plumbing for AI coding agents

> The Toolkit connects Cursor, Claude Code, Codex and similar IDEs to a CloudBase environment through a plugin, Agent Skills and an MCP server. It is useful when generated frontend code needs a real database, auth and functions behind it.

**TencentCloudBase/CloudBase-AI-Toolkit** — Backend for AI coding agents on CloudBase — database, auth, functions via Plugin, Skills & MCP.

- Repository: https://github.com/TencentCloudBase/CloudBase-AI-Toolkit
- Website: https://docs.cloudbase.net/ai/cloudbase-ai-toolkit/
- Stars: 1,129 · Forks: 146
- Language: TypeScript
- License: MIT
- Published: 2026-09-10 · Updated: 2026-09-10 · Language: en
- Canonical page: https://hysenlabs.com/projects/tencentcloudbase-cloudbase-ai-toolkit

## The gap between generated code and a running backend

AI IDEs are good at producing application code. The README states the usual blocker is everything behind it: schemas, permissions, functions, storage, environments and release. A model can write a login form in seconds and still leave you hand-configuring a database collection, an auth provider and a deploy target.

CloudBase AI Toolkit targets that gap. It is the integration layer between CloudBase, Tencent Cloud's backend that bundles database, storage, auth, cloud functions and Cloud Run, and the AI tools that write the code. The README's one-line framing is "AI writes the code. CloudBase runs the backend."

The audience is narrow and specific: developers already using a CloudBase environment who want their agent to operate it. If you are not on CloudBase, nothing here applies to you yet. The Toolkit does not create the backend for you, and the README is explicit that you still need your own CloudBase environment.

## Three pieces: Plugin, Agent Skills, MCP

The Toolkit ships as three cooperating parts, described in a table in the README.

The Plugin installs the MCP Server, the Agent Skills and the Hooks together, so you do not wire each IDE by hand. Agent Skills are scenario-specific instructions covering Web, Mini Program, database, auth and functions, aimed at steering generated code toward workable CloudBase practice rather than generic patterns. MCP is the operational layer: login, query and change data, manage functions and hosting, read logs, all from the conversation.

The split matters when you are deciding what to install. The README's guidance is to prefer the Plugin for the full stack, and to use Skills alone when you only need knowledge constraints without operational reach. That is a real distinction: Skills shape what the model writes, MCP changes what exists in your environment.

The repository layout reflects the separation. There are directories for skills, plugins, plugin specs, a dsh-plugin, a platform-kit, and an mcp package, with the npm package published as @cloudbase/cloudbase-mcp. The repo also carries per-IDE directories such as .cursor, .claude, .codebuddy and .cursor-plugin, plus instruction files like AGENTS.md, CLAUDE.md and CODEBUDDY.md.

## Installing the Toolkit and running a first real task

The README's fastest path is not a command at all. It is a prompt you paste into your AI IDE, after which the agent reads a setup document and completes the configuration:

```text
Set up CloudBase for me:
1. Open https://docs.cloudbase.net/skill.md and complete the setup following its instructions.
2. Tell me when you're done, and suggest the most relevant next step.
```

The README says the agent reads skill.md and completes the setup. You should expect the agent to report back when it is done and propose a next step.

If you prefer to drive the install yourself and your tool follows the Open Plugin Spec, the README gives one command. There is a CNB mirror as a fallback:

```bash
npx plugins add TencentCloudBase/cloudbase-plugin
```

For tools that only speak MCP, the README shows a configuration block. The server is launched through npx with the latest tag:

```json
{
  "mcpServers": {
    "cloudbase": {
      "command": "npx",
      "args": ["@cloudbase/cloudbase-mcp@latest"]
    }
  }
}
```

Claude Code and Codex take a different route: add this repository as a marketplace, then install the cloudbase plugin. CodeBuddy, WorkBuddy, ZCode and Kimi have built-in CloudBase plugins or connectors according to the README, and CodeBuddy can also install through a plugin marketplace. If you would rather keep one CLI across many tools, the CloudBase AI CLI is `npm i -g @cloudbase/cli && tcb ai`.

One warning from the README is easy to miss and worth repeating: for marketplace IDEs, use this repository, and do not also run `npx plugins add` on the same tool. Running both paths against one IDE is the setup mistake the documentation explicitly calls out.

## What the MCP surface actually touches

The release notes for v2.33.x and v2.32.x sketch the operational surface better than the README table does. Functions and Apps gained custom container-image deploy with async status query, plus a cloud upload channel through getUploadUrl and deployApp with a cosTimestamp field. Cloud API and deploy gained callCloudApi, which the notes say opens monitor and postgres services, and a declarative deploy flow through deployPlan and deployApply.

Environment binding is handled through cloudbaserc.json, which the release notes describe as a field-level fallback for envId, region and site, accepting either literals or `{{env.KEY}}` placeholders. That is the file to look at when an agent picks the wrong environment.

There is also a security-relevant default: v2.32.x introduced default env-var masking in queryFunctions and queryCloudRun. The same release added international-site login routing through `TCB_SITE=intl`. For Mini Program work, queryMessagePush and manageMessagePush handle event and message-type subscriptions with idempotent merge, and message push is aware of whether the target is a cloud function or a container, using ensureContainerMode and setContainerCallback.

Error handling is opinionated. The v2.31.x notes say CloudRun getDeployLog failures such as CODING login problems or image deploys with no build now rewrite into getProcessLog or getDeployRecords next steps instead of raw English errors. That is a deliberate choice to keep the agent on a productive path rather than surfacing a stack trace, and it cuts both ways: the agent gets better guidance, and you get less raw detail unless you go looking.

## Where the Toolkit is the wrong tool

The README states the limitation plainly: the Toolkit provides capability and path, not judgment, and you should confirm sensitive actions the AI proposes. That is not boilerplate. An agent with MCP access can change data, manage functions and hosting, and trigger deploys. If your workflow needs an approval gate between proposal and execution, that gate is yours to build, not something the Toolkit supplies.

The harder boundary is portability. This is CloudBase-specific by construction. The topics list includes postgresql and supabase, but the Toolkit's job is to drive CloudBase, so adopting it means accepting that backend as the target. If you are running your own Postgres and a self-hosted auth service, the MCP tools here have nothing to operate on.

There is also a setup complexity cost. The README lists distinct paths for marketplace IDEs, Open Plugin Spec tools, the CloudBase AI CLI, built-in IDE connectors, and MCP-only configuration. Choosing wrong is not fatal, but running two paths on one tool is the failure the documentation warns about. Budget time for reading the per-IDE setup pages rather than guessing.

Finally, the repository is a monorepo with build scripts that regenerate tool JSON, tool docs, prompts data and compatibility config. If you intend to modify the Toolkit rather than consume it, you are signing up for that build chain, not just an npm install.

## How this differs from pointing an agent at a database directly

The obvious alternative is giving your AI tool a generic database MCP server and letting it write SQL against a Postgres instance you already run. That approach is more portable and has no vendor coupling. It also hands the model raw schema and query access with none of the CloudBase-specific context.

The difference in approach is the Skills layer. A generic database MCP server exposes tables and queries. CloudBase AI Toolkit pairs MCP with scenario skills covering Web, Mini Program, database, auth and functions, plus Hooks installed by the Plugin, so the agent is steered toward CloudBase conventions while it works. The Toolkit also reaches past the database into functions, hosting, storage, logs and deploys, which a database-only server does not.

A second alternative is the CloudBase AI CLI, which the README lists for developers who prefer one CLI across many tools. It is not a competitor so much as a different door into the same backend, and the choice comes down to whether you want the agent inside your IDE or a terminal-first workflow.

If you want a plugin-based install without CloudBase, the repository's related projects list includes cloudbase-sites-plugin for Vite Web create and deploy, and a separate skills catalog installable with `npx skills add tencentcloudbase/skills --skill <name>`. Those stay inside the CloudBase orbit.

## Maintenance, licence and what upgrades cost

The repository is not archived, and the last push was on 2026-09-10. Releases have been frequent: v2.33.2 and v2.33.1 both landed on 2026-09-08, with v2.33.0 on 2026-09-04. The changelog is the place to read before upgrading, because the recent entries are behavioural changes rather than additions, such as error rewriting, env-var masking defaults, and login routing for the international site.

The licence is MIT, which is permissive and places few obligations on how you use or redistribute the code. That is a statement about the licence text, not legal advice; if you are embedding the Toolkit in a product, have your own counsel review it.

Upgrade cost is the part worth planning. The MCP server is consumed through `@cloudbase/cloudbase-mcp@latest`, so an unpinned configuration pulls whatever is current when the process starts. Given the pace of releases and the behavioural changes in the v2.31 through v2.33 notes, pinning a version in your MCP config and reading the changelog before moving is the lower-surprise path. The repository also ships lefthook.yml, renovate.json and a set of check scripts for README sync, i18n coverage and compatibility diffs, which tells you the maintainers treat drift between the docs, the skills and the tool definitions as something to catch automatically.

## Conclusion

Adopt it if you already run a CloudBase environment and want your AI IDE to create collections, deploy functions and read logs without leaving the chat. Skip it if you have no CloudBase account or if you need the agent to make permission decisions on its own, because the README states the Toolkit provides capability and path, not judgment. Before wiring anything up, confirm which install path your IDE expects, since the README warns against running the plugin installer and the marketplace route on the same tool.

## FAQ

### What is CloudBase AI Toolkit?

It is the integration layer that connects AI coding tools to CloudBase, Tencent Cloud's backend offering database, storage, auth, cloud functions and Cloud Run. It ships an MCP server published as @cloudbase/cloudbase-mcp, a set of Agent Skills, and plugins for IDEs such as Cursor, Claude Code and Codex.

### How do I install CloudBase AI Toolkit?

The README's fastest route is pasting a setup prompt into your AI IDE so the agent reads https://docs.cloudbase.net/skill.md and completes configuration. Otherwise you add this repository as a marketplace for Claude Code or Codex, run npx plugins add TencentCloudBase/cloudbase-plugin for Open Plugin Spec tools, or configure the MCP server directly.

### Can I use CloudBase AI Toolkit without a CloudBase environment?

No. The README states you still need your own CloudBase environment, and the Toolkit exists to operate that backend rather than create it.

### Does CloudBase AI Toolkit decide which actions are safe to run?

No. The README says the Toolkit provides capability and path, not judgment, and that you should confirm sensitive actions the AI proposes.

## Sources

- [License: MIT](https://github.com/TencentCloudBase/CloudBase-AI-Toolkit/blob/main/LICENSE)
- [Project website](https://docs.cloudbase.net/ai/cloudbase-ai-toolkit/)
- [README](https://github.com/TencentCloudBase/CloudBase-AI-Toolkit/blob/main/README.md)
- [Releases](https://github.com/TencentCloudBase/CloudBase-AI-Toolkit/releases)
- [TencentCloudBase/CloudBase-AI-Toolkit on GitHub](https://github.com/TencentCloudBase/CloudBase-AI-Toolkit)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/tencentcloudbase-cloudbase-ai-toolkit