03 · MCP and plugins

What MCP is

MCP (Model Context Protocol) wires external tools and services into WorkBuddy, giving the AI extended abilities: sending notifications, connecting to business systems, reaching data sources.

The official analogy is "USB for AI" — just as a computer connects peripherals over USB, MCP connects the AI to external tools and data sources.

WorkBuddy has MCP configuration built into the UI. No coding, no hand-editing config files — it is all visual.

What it gives you

Capability	Description
Context sharing	Supply file contents, database records and business information to the model
Tool invocation	Expose file operations, API calls and message sending to the model
Composable workflows	Chain several tools and services into an automated flow
Data control	Run locally or in a controlled way, balancing flexibility with security

For ready-made MCP services, the Tencent Cloud MCP marketplace is the official entry point into the wider ecosystem.

Two configuration scopes

Scope	For	Config file
User	Configure once, reuse everywhere	~/.workbuddy/mcp.json
Project	This project only, isolated	<project>/.workbuddy/mcp.json

How to choose:

• Capabilities you use across projects → user scope. WeCom bot notifications, a shared knowledge base — configure once, available everywhere.
• Services specific to one project → project scope. Client A's CRM API has no business appearing in client B's project.

A complete example: the WeCom bot

The official starter integration, and a good one — the path is clear and the result is immediately visible.

1) Get a webhook URL

In a WeCom group: add a group bot → get the webhook URL. It looks like:

https://qyapi.weixin.qq.com/cgi-bin/webhook/send?key=xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx

2) Open the MCP configuration

Sidebar Plugins (插件) → MCP servers (MCP 服务器) at the top right → Configure MCP (配置 MCP).

3) Write mcp.json

Paste this into the editor, replacing your-webhook-url with the real address, and save:

{
  "mcpServers": {
    "wecom": {
      "command": "uvx",
      "args": ["wecom-bot-mcp-server"],
      "env": {
        "WECOM_WEBHOOK_URL": "your-webhook-url"
      }
    }
  }
}

4) Check the status

After saving, look at the server's indicator:

Status	Meaning
🟢 Green	Connected and usable
🔴 Red	Configuration problem — check the config, the command environment, or the address

5) Use it

Just describe what you want; WorkBuddy calls the right MCP server:

Notify the WeCom group: the production release is live, please @Zhang San to verify.

Where it fits

Scenario	Description
Release notifications	Tell the project group automatically when a version ships
Task reminders	Push to-dos, sign-offs and hand-offs into a WeCom group
Process integration	Connect WorkBuddy's task handling to external platforms
Capability extension	Reach richer external services through more MCP servers

Best practice

Configuration

• Shared capability, user scope — configure notification capability once and reuse it
• Project-specific, project scope — keep them isolated
• Start from a proven example — get WeCom Bot working before attempting anything exotic
• Be explicit in your description — say who to notify, what to say, whether to @ anyone; invocation is far more stable

Security

• Guard the webhook URL — it is the bot's credential; anyone with it can post to your group
• Check the JSON — malformed brackets, quotes and commas are the single most common failure
• Watch the indicator — green works, red needs investigating

Plugins: a bundle of capabilities

A plugin (插件) packages a related set of capabilities for distribution. One plugin can contain Skills, external tool connections, slash commands and hooks, all usable after a single install.

What a plugin contains

Component	Description	Example
Skills	How to accomplish a class of task	—
MCP connections	Reach external services	Connect mail, cloud drives
Slash commands	Trigger a function with a short command	Type /weekly to generate a weekly report
Hooks	Run automatically at a specific moment	Check formatting after a file is generated
Agents	A dedicated AI assistant for a specific task	—

Different plugins carry different numbers of components, but installation and use are identical.

The plugin market

Browse recommended plugins and install with one click. Installed plugins can be enabled, disabled or updated; click a card to manage it individually. Keyword search is supported.

Plugin or Skill?

Need one capability → install a Skill. Need a set that works together (say, "weekly reports" needs a Skill, a slash command and a schedule) → look for a plugin.

A plugin's value is not having to assemble it yourself; the cost is accepting all its components at once. Read the component list before installing.

Custom models

Worth covering here because, like MCP, it is about wiring something external in.

Manage custom models under Settings → Models through the GUI — no config file editing. Configuration persists automatically, and choosing a standard provider fills in capability flags (tool calling, image input) for you.

Three routes

By provider. Pick from the list and the URL, model list and capability flags fill in automatically; you supply the API key. Supports Tencent Cloud Token Plan (several personal and enterprise tiers), various Coding Plans, and presets for the major vendors.

Local deployment. Connect Ollama — an open-source local model runner. After installation one command pulls and runs an open model; it listens on port 11434 by default and exposes an OpenAI-compatible interface. Three good reasons:

• Privacy and internal-network compliance: code and conversation never leave the machine
• Zero-cost experimentation: no tokens, no paid API key
• Offline: works on a plane, on an air-gapped dev box, anywhere without a network

Custom. Service not on the list — enter URL, API key and model name yourself.

Custom protocol

When a model service uses a non-standard URL path (behind a gateway or proxy layer), turn on the custom protocol switch in advanced settings:

State	Behaviour
Off (default)	Uses the standard /chat/completions path, validating and completing the address
On	Requests the exact URL you entered, skipping validation and path completion

Costs do not go through credits

All costs from a custom model — tokens, subscriptions — are paid by you to that third party. WorkBuddy acts only as a transport, forwarding your input and returning the model's output.

Usage can trigger model calls continuously, so watch your third-party account balance — that spend never appears on your WorkBuddy credit statement.

Old configuration still works

Custom models configured through ~/.codebuddy/models.json continue to work after the UI upgrade, and can be viewed, edited or deleted through the interface without touching the file.

FAQ

The MCP server stays red.

In order: JSON syntax (brackets, quotes, commas) → is the runtime named in command installed (uvx, npx) → are the URL and credentials right → can you reach the target address from this network.

Is MCP the same as a connector?

A connector is a product-level concept, and some official connectors are MCP implementations underneath. A custom connector is essentially "bring your own MCP server". Think of connectors as MCP that has already been configured for you.

Does project-scope configuration sync to team members?

The project-scope mcp.json lives in the project directory and travels with it. But project connector authorisation splits into public and personal, with different credential handling — see Projects and teamwork.

Next

04 Experts and expert teams — give it a professional identity.