# opencode-mcp-lab

Learning MCP (Model Context Protocol) by building my own servers and using them with opencode + DeepSeek.

## Servers

| Server | What it does |
|---|---|
| `opencode-hello-mcp` | One tool, `greet-tool`: the "hello world" of MCP |

## Setup

```bash
npm install
```

## Test a server without AI (free)

```bash
npm run inspect
```

## Use it in opencode

```bash
opencode
```

Then ask: `greet Vivek using the greet tool`

## Connecting & disconnecting MCP servers

MCP servers are wired up in `opencode.json` (project root). opencode reads it on startup, so **restart opencode after any change** (`/exit`, then `opencode`).

### Connect a server

Add an entry under `"mcp"` with `"enabled": true`:

```json
{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "opencode-hello-mcp": {
      "type": "local",
      "command": ["node", "servers/hello-mcp/index.ts"],
      "enabled": true
    }
  }
}
```

- `"opencode-hello-mcp"`: the name opencode shows (tools appear as `opencode-hello-mcp_greet-tool`)
- `"type": "local"`: opencode starts the server on your machine
- `"command"`: how to start it (run from the project root)

### Check it's connected

```bash
opencode mcp list
```

Or inside opencode: the status bar shows `1 MCP`, and `/status` shows whether each server is connected.

### Disconnect: 3 ways

**1. Turn it off temporarily (recommended)**: set `"enabled": false` in `opencode.json` and restart opencode:

```json
"opencode-hello-mcp": {
  "type": "local",
  "command": ["node", "servers/hello-mcp/index.ts"],
  "enabled": false
}
```

Set it back to `true` to reconnect.

**2. Toggle it for this session only**: inside opencode, type `/mcps` and switch servers on or off. No file changes; resets next time you start opencode.

**3. Remove it completely**: delete the whole `"opencode-hello-mcp": { ... }` block from `opencode.json` and restart opencode.

> Disconnecting never deletes your server code. `servers/hello-mcp/index.ts` stays, so you can reconnect anytime.

### Why disconnect?

Every connected server's tool descriptions are sent to the AI with **every message**, which costs tokens. Keep only the servers you're using switched on.