MCP Tools

After reading this page you will be able to: add and manage MCP tools so the AI can call external services plus built-in image, browser, and desktop-automation capabilities.

MCP (Model Context Protocol) is the standard way to plug external tools into an AI: configure one “MCP server” and the AI gains a set of tools it can call — querying databases, driving a browser, generating images, and more.

1. Opening tool settings

Where: Settings → Tools

The upper section, “MCP Tools Configuration”, lists all MCP servers; the lower section holds the built-in capability switches (see section 5). When the list is empty you will see “No MCP servers found”.

2. Adding MCP servers

Click Add MCP in the top-right corner. Three ways:

Way	Who it is for	How
Paste config	You have a ready config	Paste the MCP JSON from the service documentation and save
Import from JSON	Same, for bulk	Paste JSON containing several servers to import them at once
One-Click Import	CLI tools such as Claude Code or Gemini CLI are installed	See section 4, scanning import

You can also pick Add via chat and let the Butler guide you.

When pasting JSON, mind the on-page hint: “Paste the MCP JSON configuration from the service documentation below. If the configuration uses commands like npx or uvx, make sure they are installed and available on your machine.” Supported transports:

Type	Trait	Example
stdio	A local command-line program	started via `command` + `args`
http / sse	A remote service address	`url` + headers

After saving, a connectivity check runs automatically — see the next section for statuses.

3. Managing and statuses

Each MCP server is one row; click to expand its configuration and tool list:

Action	Notes
Edit	Change name fields and config; the top-level server name cannot be changed while editing
Delete	Asks “Are you sure you want to delete this MCP server?” first
Check MCP Availability	The refresh button on the row; success shows “Connection test successful”
Login	OAuth-capable remote services show a “Login” button; success shows “OAuth login successful”

Status icons (hover for details and check time):

Status	Meaning
Manual check passed	Last check succeeded, with check time
Authenticated	OAuth signed in
Not tested	No check yet
Failed	Last check failed, with check time and the likely cause

4. Scanning import (One-Click Import)

If CLI agents (Claude Code, Gemini CLI, …) are installed, click One-Click Import next to Add MCP and follow the three-step wizard:

Step	What the page does
1 Select Agent	A dropdown lists the detected CLIs; pick one
2 Fetch mcp	Shows “Detected N MCP servers”, marking each as Will import / Skipped
3 Import Success	Shows “Imported N MCP servers”

Skipped entries state the reason; common ones: Already imported, Unsupported, Managed by a CLI plugin and can’t be imported here, Disabled in the CLI, Requires authentication in the CLI first, Unavailable in the CLI right now.

5. Built-in MCP tools

These three ship out of the box and are managed on the Settings → Tools page:

Built-in tool	What it is	How to use
Image Generation	A built-in MCP server; pick an “Image Model” first	Turn on the switch; when the Image Model dropdown has no options, click “Go to configure” to set one up on the Model page
In-app browser	Enabled by default. The agent opens pages, clicks, and fills forms in the in-app browser side panel; it can only drive that one view — it cannot reach the AIPC UP interface and never opens a separate Chrome window	Just chat with your agent; “Let the agent use the in-app browser” turns access off entirely; “Clear browsing data” signs out of every site and clears the cache, which cannot be undone
Desktop Automation MCP	Lets the AI call the `desktop_automation_*` tool family to operate the Windows keyboard and mouse	See below

The in-app browser shares the login state of the browser you see in AIPC UP, so the agent can pick up pages you are already signed into; use “Clear browsing data” — with its confirmation — to wipe cookies and cache.

Image Generation supports these image models (hovering the Image Model field shows the same note): Google Gemini / Vertex AI image models (e.g. gemini-2.5-flash-image), OpenRouter chat-style image models (e.g. nano-banana, gemini-image), and AntigravityTools (gemini-3-pro-image); models like gpt-image-* / dall-e-* / Imagen / FLUX use a different API and are not yet supported.

6. Attaching to assistants and common errors

Attach to assistants: in the assistant details under Default configuration, pick MCP in Assistant Center, or jump there via “Go to MCP settings”; sessions can also select tools temporarily — see Permissions and Safety
Startup errors: when a server fails to start, its status tells you what is missing and what to install; fix it and check again. The “Status message” column below is abbreviated — the in-app message is the source of truth. It may say, for example: “Command bun was not found. Install Bun (which includes bun/bunx), restart AIPC UP, or configure this MCP server to use an absolute command path.” The exact wording can vary by app version.

Status message	Fix
Command bun was not found. Install Bun	Install Bun, restart the app, check again
Command uv was not found. Install uv	Install uv, restart the app, check again
Command python was not found. Install Python	Install Python, restart the app, check again
Other “command not found”	Install the command, or point this MCP server to an absolute command path
Check timed out / could not reach	Check whether the service needs network or a proxy, then test again

How to tell it worked

Your server appears in the MCP list on Settings → Tools with status “Manual check passed”
After signing in, OAuth services show “Authenticated”
The built-in Image Generation switch turns on and the Image Model dropdown offers models
The AI actually calls your tools in a conversation — for example, driving the in-app browser while you watch the side panel