# Your desktop AI should call tools — not just talk about them

Cloud chatbots are great at explaining an API. They are usually terrible at *using* the one that lives on your laptop, in your CRM, or behind a private database. You end up copy-pasting outputs between tabs: search here, paste there, ask the model to “use this JSON,” then lose the thread when the browser tab dies.

Makers who actually ship marketing and research workflows need the opposite: an agent that can discover tools, call them with structured arguments, and keep that loop on a machine they control.

That is what **MCP Tools** in [aiFetchly](https://www.aifetchly.com) is for.

## What MCP adds to a local-first desktop agent

[MCP (Model Context Protocol)](https://docs.aifetchly.com/docs/ai-outreach/mcp-tools) is an open standard for wiring external tools into an AI session. In aiFetchly, MCP servers show up inside the AI Marketing Assistant chat (open chat with `Cmd/Ctrl + K`, then the MCP Tools button). Each server exposes tools the agent can call during a conversation — web search, database queries, custom APIs, CRM hooks, analytics pulls, whatever you attach.

You are not locked into one cloud workspace. The agent runs on your desktop; the tools are the ones *you* configure.

## How you add a server (two modes)

1. Open the MCP Tools dialog from chat.
2. Click **Add Server**.
3. Choose **Form Mode** (fields for name, transport, host/command, auth) or **JSON Mode** (paste a Claude Desktop–style `mcpServers` block).
4. Save, then **Discover Tools** so aiFetchly pulls the live tool list.
5. Toggle individual tools on or off if a server ships more than you want exposed.

JSON mode accepts the familiar shape:

```json
{
  "mcpServers": {
    "server-name": {
      "command": "uvx",
      "args": ["package-name"]
    }
  }
}
```

Transports include **Stdio** (local processes / `npx` / `uvx`), **SSE**, and **WebSocket**. Auth can be none, API key, bearer token, or custom. You can also bundle MCP servers inside a **plugin** and manage them from the Plugin Manager — same tools, less one-off config.

## Why this matters for real workflows

- **Research with tools, not tabs.** Point the agent at a search or scrape MCP and keep the conversation grounded in live results instead of stale memory.
- **Private data stays private.** Database and internal API MCPs talk from *your* machine under *your* credentials — not a shared cloud sandbox you do not control.
- **Composable with the rest of aiFetchly.** MCP pairs naturally with Knowledge Library grounding, Skills, Subagents, and permission-gated hooks when you want tool use plus approval boundaries.
- **Standards-friendly.** Claude Desktop–compatible JSON means configs you already have often drop in with small edits.

Docs walk through discovery, enable/disable, connection tests, and tool naming (`mcp_` prefixes in chat): [MCP Tools](https://docs.aifetchly.com/docs/ai-outreach/mcp-tools).

## A practical starting recipe

1. Install aiFetchly and confirm AI provider settings work: [installation](https://docs.aifetchly.com/docs/getting-started/installation).
2. Open AI chat → MCP Tools → Add Server.
3. Start with one low-risk Stdio server you already trust (or paste a known-good JSON example via Load Example).
4. Discover tools, enable only the ones you need, run a Test Connection.
5. Ask the agent to use a specific tool by name in a real task — lead research, outreach prep, or a local query — and watch the call succeed before stacking more servers.

If you already use plugins, check whether an MCP server arrived bundled; plugin-owned servers appear with an owner-plugin label so you do not re-add them by hand.

## Bottom line

An agent that cannot call tools is a very polite search engine. An agent that can attach MCP servers on your desktop becomes a workstation: search, query, draft, and act with the integrations *you* chose.

Feature overview: [aifetchly.com/features](https://www.aifetchly.com/features). Intro docs: [getting started](https://docs.aifetchly.com/docs/getting-started/introduction). Source: [github.com/robertzengcn/aiFetchly](https://github.com/robertzengcn/aiFetchly).

- Site: https://www.aifetchly.com
- Docs: https://docs.aifetchly.com/docs/ai-outreach/mcp-tools
- Features: https://www.aifetchly.com/features
- GitHub: https://github.com/robertzengcn/aiFetchly

*Posted as community notes for makers evaluating local-first desktop AI agents — not a directory listing.*