cloudbase-mcp

Imported from tencentcloudbase/skills (skills/cloudbase-mcp/SKILL.md). Install upstream with npx skills add tencentcloudbase/skills --skill cloudbase-mcp. Copyright stays with the author (MIT).

CloudBase MCP

Open source:

• GitHub: https://github.com/TencentCloudBase/CloudBase-AI-ToolKit
• Domestic mirror: https://cnb.cool/tencent/cloud/cloudbase/CloudBase-AI-ToolKit

The server source is mcp/ in that repo. The package name is @cloudbase/cloudbase-mcp.

Sibling skills (local only)

Sibling CloudBase skills ship beside this skill. Use local relative paths such as ../cloudbase-cli/SKILL.md.

If a referenced sibling skill file is missing from this environment, ask the user to install the full CloudBase plugin (or the missing skill). Do not HTTP-fetch remote skill or protocol markdown into the agent context.

• MCP setup, mcporter commands, and auth examples: ../cloudbase/references/mcp-setup.md
• CLI fallback when MCP tools are not loaded: ../cloudbase-cli/SKILL.md
• MCP vs CLI decision tree: ../cloudbase/references/tooling-fallback.md

When to use

• The user is connecting MCP, or choosing local versus remote.
• The user or the agent calls CloudBase tools from the shell with mcporter, or needs a tool schema before calling.
• The user needs the site, region, cloud-mode, or plugin switches that already exist.
• A bug is in this repo's MCP server, skills, or CLI, and the fix should be a pull request.
• A CloudBase tool failed and the user needs to report it, but cannot share the whole session. Read references/issue-report.md first and build the report before answering anything else.

Do not use this for

• Creating or changing cloud resources. Use the matching domain skill, or ../cloudbase-cli/SKILL.md when MCP tools are not loaded.
• Cloud platform, permission, or account problems that are not code in this repo. Those go to an issue, not a pull request.

Connection

The hostname decides the site. Remote URLs do not take a site query parameter.

• Interactive clients: put only the URL in the IDE config. The IDE opens a browser for login. This is the usual path.
• CI: add env_id and headers X-TencentCloud-SecretId / X-TencentCloud-SecretKey.
• Local stdio for the international site: TCB_SITE=intl and TCB_REGION=ap-singapore.

Pick the mode by these rules:

• Remote — the default. Interactive sign-in (browser OAuth) plus anything that runs entirely in the cloud.
• Local — when any one of these applies:
  • the account signs in with a CloudBase API key (CLOUDBASE_API_KEY); remote has no API-key path, it accepts OAuth or Tencent Cloud SecretId/SecretKey only
  • one session has to cover both the domestic and the international site; a remote URL is bound to a single site, while local switches with TCB_SITE
  • the task needs a tool that remote does not register: manageStorage for uploading or downloading local files, downloadTemplate, manageCloudRun, or the declarative deployBuild / deployPlan / deployApply set

Deploying functions or apps does not require local mode. Remote deploys through the two-step upload: queryFunctions(action="getFunctionUploadUrl") or queryApps(action="getUploadUrl") → PUT the zip → manageFunctions with the code triple or manageApps(action="deployApp", cosTimestamp=…). What remote lacks is the declarative cloudbaserc executor and a dry-run before applying.

The international site does not expose NoSQL tools in either mode.

Readiness check: missing tools are not the same as unconfigured

Run this before concluding MCP is unavailable, and before reaching for any workaround. Four different failures look alike from the outside and need different fixes.

1. Wiring — is a tool registered in this session? If no cloudbase-prefixed MCP tool appears in the current session's tool list, treat it as a wiring problem, not a credential problem. In hosts with a connector manager (WorkBuddy and similar), check the installed connector list first and reconnect from there. A bound connector is deliberately excluded from search and recommendation results, so "search returns nothing" does not mean "not installed". Search with the bare entity name (cloudbase), not a sentence — sentence-shaped queries match unrelated entries, and keyword filters do not help. Install only when it is genuinely absent.
2. Credential — can the tool actually reach the environment? When the tools are present, call auth(action="status") and require auth_status: "READY" plus the expected current_env_id before any cloud write. Anything else is a credential or binding problem, not a wiring problem.
3. Session lifetime — when is a restart required? Only after credentials change (first authorization, environment switch, key rotation); the running session caches the credential it started with. Enabling, disabling, or unbinding a connector does not require a new session — and unbinding does not revoke calls already flowing through an open session. To cut access immediately, close the session and rotate the credential.
4. Triage instead of guessing. No tool or tool not found → wiring. auth not READY → credential. READY but wrong current_env_id → binding. Network, signature, or permission errors → credential or gateway; reinstalling the connector will not help.

Which call fixes which state. The four cases above map to four different calls. Run the check; do not jump to a reinstall.

Credentials live in ~/.config/.cloudbase/auth.json (mode 600); authSource records how they were obtained, and a connector reads the same file through login_by_api_key. Do not write a key into mcp.json: that copy drifts from the CLI and the two disagree about which environment is current.

Fallback is not a workaround. If MCP tools are missing, configure them for the next session and use the tcb CLI now (../cloudbase-cli/SKILL.md). Do not use an application's data-plane SDK or API key to perform a control-plane operation (create tables, deploy, bind domains) — it fails with errors that point at the wrong cause, and the detour costs more than the fallback.

Calling tools with mcporter

When this session has no native MCP tools, use mcporter. The command list and the auth examples live in ../cloudbase/references/mcp-setup.md. Read that file before login, environment binding, or manageEnv.

Check whether the server is configured:

npx mcporter list | grep cloudbase

If it is missing, add a stdio server named cloudbase in config/mcporter.json:

{
  "mcpServers": {
    "cloudbase": {
      "command": "npx",
      "args": ["@cloudbase/cloudbase-mcp@latest"],
      "description": "CloudBase MCP",
      "lifecycle": "keep-alive"
    }
  }
}

Read the schema before the first call. describe without --all-parameters omits parameters.

npx mcporter describe cloudbase --all-parameters
npx mcporter list cloudbase --schema

Call a tool with server.tool and key=value arguments:

npx mcporter call cloudbase.queryEnv action=list --output json

Do not put a Secret ID, Secret Key, or environment ID into mcporter.json. Login stays on cloudbase.auth or tcb login, as written in mcp-setup.md.

Environment

Plugin names are the ones in mcp/src/server.ts (env, database, functions, hosting, storage, and the rest). A local server and a hosted remote server enable the same default set; turn one off for a given caller with disable_plugins on the remote URL or CLOUDBASE_MCP_PLUGINS_DISABLED on a local server.

Contributing

Read CONTRIBUTING.md at the repo root. Fork, use a feature branch, and follow the commit types listed there (feat, fix, docs, and the rest). Install with corepack and pnpm install.

• A bug in MCP, a skill, or the CLI: open a pull request.
• The cloud platform, a permission, or the account: open an issue. Do not put an environment ID or a secret in the text.
• prepareFeedback only returns the new-issue page for the current site. Write the text in the chat. The user logs in and submits it.
• The user cannot or will not share the conversation: build the report first, following references/issue-report.md. It has two parts — the call chain (tools, arguments, error wording) and the intent review (plan vs actual, which source pulled you off, the first step that went wrong). Replace the customer's business values — table names, entities, data rows — and mark every field you could not observe as unknown.